Spaces:
Running
Running
File size: 108,808 Bytes
5504ed3 6fbe974 5504ed3 6fbe974 5504ed3 6fbe974 5504ed3 6fbe974 5504ed3 6fbe974 5504ed3 6fbe974 5504ed3 6fbe974 5504ed3 96f4107 5504ed3 6fbe974 5504ed3 0106486 6fbe974 0106486 5504ed3 6fbe974 0106486 5504ed3 0106486 5504ed3 6fbe974 5504ed3 6fbe974 0106486 5504ed3 6fbe974 5504ed3 6fbe974 5504ed3 6fbe974 5504ed3 6fbe974 5504ed3 6fbe974 5504ed3 6fbe974 5504ed3 6fbe974 5504ed3 2525fb4 6fbe974 5504ed3 6fbe974 5504ed3 6fbe974 5504ed3 6fbe974 5504ed3 6fbe974 5504ed3 6fbe974 5504ed3 6fbe974 5504ed3 6fbe974 5504ed3 cd67641 5504ed3 cd67641 5504ed3 cd67641 5504ed3 cd67641 5504ed3 cd67641 5504ed3 cd67641 5504ed3 cd67641 5504ed3 cd67641 5504ed3 6fbe974 5504ed3 6fbe974 5504ed3 6fbe974 5504ed3 6fbe974 5504ed3 49de2fb 5504ed3 6fbe974 49de2fb 5504ed3 6fbe974 5504ed3 6fbe974 5504ed3 6fbe974 5504ed3 6fbe974 5504ed3 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764 1765 1766 1767 1768 1769 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815 1816 1817 1818 1819 1820 1821 1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832 1833 1834 1835 1836 1837 1838 1839 1840 1841 1842 1843 1844 1845 1846 1847 1848 1849 1850 1851 1852 1853 1854 1855 1856 1857 1858 1859 1860 1861 1862 1863 1864 1865 1866 1867 1868 1869 1870 1871 1872 1873 1874 1875 1876 1877 1878 1879 1880 1881 1882 1883 1884 1885 1886 1887 1888 1889 1890 1891 1892 1893 1894 1895 1896 1897 1898 1899 1900 1901 1902 1903 1904 1905 1906 1907 1908 1909 1910 1911 1912 1913 1914 1915 1916 1917 1918 1919 1920 1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931 1932 1933 1934 1935 1936 1937 1938 1939 1940 1941 1942 1943 1944 1945 1946 1947 1948 1949 1950 1951 1952 1953 1954 1955 1956 1957 1958 1959 1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971 1972 1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 2025 2026 2027 2028 2029 2030 2031 2032 2033 2034 2035 2036 2037 2038 2039 2040 2041 2042 2043 2044 2045 2046 2047 2048 2049 2050 2051 2052 2053 2054 2055 2056 2057 2058 2059 2060 2061 2062 2063 2064 2065 2066 2067 2068 2069 2070 2071 2072 2073 2074 2075 2076 2077 2078 2079 2080 2081 2082 2083 2084 2085 2086 2087 2088 2089 2090 2091 2092 2093 2094 2095 2096 2097 2098 2099 2100 2101 2102 2103 2104 2105 2106 2107 2108 2109 2110 2111 2112 2113 2114 2115 2116 2117 2118 2119 2120 2121 2122 2123 2124 2125 2126 2127 2128 2129 2130 2131 2132 2133 2134 2135 2136 2137 2138 2139 2140 2141 2142 2143 2144 2145 2146 2147 2148 2149 2150 2151 2152 2153 2154 2155 2156 2157 2158 2159 2160 2161 2162 2163 2164 2165 2166 2167 2168 2169 2170 2171 2172 2173 2174 2175 2176 2177 2178 2179 2180 2181 2182 2183 2184 2185 2186 2187 2188 2189 2190 2191 2192 2193 | from __future__ import annotations
from collections import defaultdict
from datetime import date, datetime, time, timedelta
import hashlib
import math
import random
import traceback
from statistics import mean, stdev
from uuid import uuid4
import pandas as pd
from sqlalchemy import desc, func, select
from sqlalchemy.orm import Session
from app.core.config import get_settings
from app.models import (
Asset,
FundamentalSnapshot,
FeedbackLoopAudit,
HistoricalPrediction,
CapitalPreservationAlpha,
LearningFocusPriority,
LearningEvent,
LearningMetric,
LearningRun,
MacroSnapshot,
MissedWinner,
MistakeAnalysis,
ModelVersion,
NewsArticle,
NewsAssetLink,
PredictionOutcome,
PriceHistory,
SignalPerformance,
StrategyMemory,
TradingGameTrade,
)
from app.services.technical_analysis_engine import TechnicalAnalysisEngine
settings = get_settings()
TIMEFRAMES = {
"short": {"horizon_days": 20, "label": "5-20 trading days"},
"mid": {"horizon_days": 63, "label": "1-3 months"},
"long": {"horizon_days": 252, "label": "6-12 months"},
}
BASE_SIGNAL_WEIGHTS = {
"trend_structure": 0.18,
"momentum": 0.16,
"volume_confirmation": 0.13,
"volatility_control": 0.12,
"support_resistance": 0.10,
"sentiment": 0.10,
"narrative": 0.08,
"fundamentals": 0.08,
"regime": 0.05,
}
MIN_MODEL_VERSION_OUTCOMES = 30
MIN_MODEL_VERSION_SIGNAL_ROWS = 3
MIN_MODEL_VERSION_SIGNAL_SAMPLE = 3
ERROR_TYPES = {
"technical_false_breakout",
"overbought_signal_ignored",
"weak_volume_confirmation",
"sentiment_overestimated",
"fundamentals_ignored",
"macro_regime_changed",
"earnings_event_underestimated",
"news_shock",
"sector_rotation",
"volatility_expansion",
"support_resistance_wrong",
"timeframe_mismatch",
"overconfidence",
"underconfidence",
"insufficient_data",
"random_market_noise",
}
class LearningLoopService:
"""Runs point-in-time historical simulations and updates Blum's strategy memory."""
def __init__(self) -> None:
self.sampler = HistoricalSamplerService()
self.point_in_time = PointInTimeDataService()
self.predictor = PredictionEngine()
self.evaluator = OutcomeEvaluator()
self.mistakes = MistakeAnalyzer()
self.memory = StrategyMemoryService()
self.model_scores = ModelScoreService()
def run_batch(
self,
db: Session,
batch_size: int | None = None,
trigger: str = "manual",
sniper_simulation_limit: int | None = None,
) -> dict:
requested_batch = int(batch_size or settings.learning_batch_size)
configured_batch = max(1, min(requested_batch, settings.learning_batch_size, 500))
daily_guard = self.daily_guard(db, configured_batch)
batch = max(1, int(daily_guard.get("effective_batch", configured_batch) or configured_batch))
run_id = f"learn-{datetime.utcnow().strftime('%Y%m%d%H%M%S')}-{uuid4().hex[:8]}"
if not daily_guard["allowed"]:
db.add(
LearningEvent(
event_type="blum_learning_loop_budget_wait",
severity="Warning",
title="BLUM Learning Loop waiting for daily budget",
description=daily_guard["reason"],
payload={"run_id": run_id, "guard": daily_guard, "policy": "Budget guards are events, not zero-output training experiments."},
)
)
db.commit()
return {"status": "budget_wait", "run_id": run_id, "guard": daily_guard}
run = LearningRun(
run_id=run_id,
trigger=trigger,
status="running",
evaluation_mode=settings.learning_evaluation_mode,
asset_universe=settings.learning_asset_universe,
batch_size=batch,
anti_overfitting_report=daily_guard,
)
db.add(run)
db.flush()
reports: list[dict] = []
errors: list[dict] = []
seen_samples: set[tuple[str, str]] = set()
for _ in range(batch):
try:
sample = self.sampler.blended_sample(db, seen_samples=seen_samples, trigger=trigger)
if not sample:
errors.append({"stage": "sample", "error": "No eligible historical sample found."})
continue
seen_samples.add((sample["asset"].ticker, sample["analysis_date"].isoformat()))
reports.append(self.run_single_sample(db, run, sample))
except Exception as exc:
errors.append({"stage": "sample", "error": f"{type(exc).__name__}: {exc}", "traceback": traceback.format_exc(limit=4)})
sniper_learning = {"status": "skipped", "reason": "No reports created."}
if reports:
try:
limit = sniper_simulation_limit if sniper_simulation_limit is not None else min(300, max(60, len(reports) * len(TIMEFRAMES) * 6))
if limit > 0:
from app.services.market_sniper import MarketSniperEngine
sniper_learning = MarketSniperEngine().simulate(db, limit=limit)
else:
sniper_learning = {
"status": "deferred",
"reason": "Sniper R-multiple simulation is deferred for this bounded learning lane.",
}
except Exception as exc:
sniper_learning = {"status": "degraded", "error": f"{type(exc).__name__}: {exc}"}
metrics = LearningDashboardService().aggregate_metrics(db)
model_version = self.model_scores.recalculate(db)
anti_overfitting = self.model_scores.anti_overfitting_report(db)
run.status = "ok" if reports else "degraded"
run.completed_at = datetime.utcnow()
run.predictions_created = len(reports)
run.outcomes_evaluated = sum(len(report.get("actual_outcome", {})) for report in reports)
run.mistakes_found = sum(len(report.get("mistakes", [])) for report in reports)
run.memory_updates = sum(len(report.get("memory_updates", [])) for report in reports)
run.summary = {
"reports_created": len(reports),
"errors": errors[:8],
"latest_reports": reports[-5:],
"dashboard_metrics": metrics,
"model_version": model_version,
"learning_mode": "alpha_loss_replay" if trigger == "alpha_loss_replay" else "random_point_in_time",
"market_sniper_learning": sniper_learning,
}
run.anti_overfitting_report = anti_overfitting
run.error_payload = {"errors": errors}
db.add(
LearningEvent(
event_type="blum_learning_loop",
severity="Info" if reports else "Warning",
title="BLUM Learning Loop completed",
description="Point-in-time historical simulations updated prediction outcomes, mistakes, strategy memory, adaptive signal scores and Market Sniper R-multiple memory.",
payload={"run_id": run_id, "reports_created": len(reports), "errors": errors[:8], "anti_overfitting": anti_overfitting, "market_sniper_learning": sniper_learning},
)
)
db.commit()
return {
"status": run.status,
"run_id": run_id,
"batch_size": batch,
"requested_batch_size": requested_batch,
"daily_guard": daily_guard,
"reports_created": len(reports),
"errors": errors[:8],
"metrics": metrics,
"model_version": model_version,
"learning_mode": "alpha_loss_replay" if trigger == "alpha_loss_replay" else "random_point_in_time",
"market_sniper_learning": sniper_learning,
"anti_overfitting": anti_overfitting,
"disclaimer": "Research learning loop only. It improves calibration and robustness; it does not create guaranteed market predictions.",
}
def run_single_sample(self, db: Session, run: LearningRun, sample: dict) -> dict:
context = self.point_in_time.context_for(db, sample["asset"], sample["analysis_date"])
prediction_context = dict(context)
prediction_context.pop("future_prices", None)
sample_metadata = dict(sample)
sample_metadata["run_trigger"] = run.trigger
sample_metadata["evaluation_mode"] = run.evaluation_mode
sample_metadata["learning_mode_metadata"] = learning_mode_metadata(run.trigger, sample_metadata)
prediction_payload = self.predictor.predict(prediction_context, db=db, sample_metadata=sample_metadata)
feedback_payload = prediction_payload.get("feedback_loop", {})
prediction = HistoricalPrediction(
learning_run_id=run.id,
asset_id=sample["asset"].id,
ticker=sample["asset"].ticker,
asset_type=sample["asset"].asset_type,
sector=sample["asset"].sector or "Unknown",
market=sample["asset"].country or "Unknown",
market_regime=context["market_context"]["market_regime"],
volatility_regime=context["market_context"]["volatility_regime"],
analysis_date=sample["analysis_date"],
initial_price=context["initial_price"],
prediction_payload=prediction_payload,
point_in_time_context=json_safe(context),
expected_direction=prediction_payload["prediction"]["dominant_direction"],
confidence=prediction_payload["prediction"]["aggregate_confidence"],
model_version=feedback_payload.get("model_version_used") or "base-static",
model_version_used=feedback_payload.get("model_version_used") or "base-static",
weights_used=json_safe(feedback_payload.get("weights_used") or {}),
learning_memory_used=json_safe(feedback_payload.get("learning_memory_used") or {}),
strategy_memory_used=json_safe(feedback_payload.get("strategy_memory_used") or {}),
research_priority_used=json_safe(feedback_payload.get("research_priority_used") or {}),
data_quality_score=context["data_quality_score"],
)
db.add(prediction)
db.flush()
outcomes = []
mistakes = []
memory_updates = []
for timeframe, frame_prediction in prediction_payload["timeframes"].items():
outcome_payload = self.evaluator.evaluate(context, timeframe, frame_prediction)
outcome = PredictionOutcome(
prediction_id=prediction.id,
ticker=prediction.ticker,
timeframe=timeframe,
horizon_days=outcome_payload["horizon_days"],
evaluation_date=outcome_payload.get("evaluation_date"),
price_at_evaluation=outcome_payload.get("price_at_evaluation"),
realized_return=outcome_payload.get("realized_return"),
max_favorable_excursion=outcome_payload.get("max_favorable_excursion"),
max_adverse_excursion=outcome_payload.get("max_adverse_excursion"),
drawdown=outcome_payload.get("drawdown"),
time_to_target=outcome_payload.get("time_to_target"),
time_to_invalidation=outcome_payload.get("time_to_invalidation"),
target_hit=bool(outcome_payload.get("target_hit")),
invalidation_hit=bool(outcome_payload.get("invalidation_hit")),
direction_correct=outcome_payload.get("direction_correct"),
false_positive=bool(outcome_payload.get("false_positive")),
false_negative=bool(outcome_payload.get("false_negative")),
missed_opportunity=bool(outcome_payload.get("missed_opportunity")),
outcome_label=outcome_payload.get("outcome_label", "inconclusive"),
confidence_calibration_error=outcome_payload.get("confidence_calibration_error"),
metrics_payload=json_safe(outcome_payload),
)
db.add(outcome)
db.flush()
outcomes.append(outcome_payload)
analysis = self.mistakes.analyze(context, prediction_payload, frame_prediction, outcome_payload)
if analysis["persist"]:
mistake = MistakeAnalysis(
prediction_id=prediction.id,
outcome_id=outcome.id,
ticker=prediction.ticker,
timeframe=timeframe,
error_type=analysis["error_type"],
severity=analysis["severity"],
predicted=json_safe(frame_prediction),
actual=json_safe(outcome_payload),
misleading_signal=analysis["misleading_signal"],
signal_to_weight_more=analysis["signal_to_weight_more"],
rule_adjustment=analysis["rule_adjustment"],
future_impact=analysis["future_impact"],
explanation=analysis["explanation"],
payload=json_safe(analysis),
)
db.add(mistake)
mistakes.append(analysis)
memory_updates.extend(self.memory.update_from_outcome(db, context, frame_prediction, outcome_payload, analysis))
self.memory.update_signal_performance(db, context, prediction_payload, outcomes)
feedback_audit = FeedbackLoopAuditService().record(db, prediction, prediction_payload, outcomes, memory_updates)
report = {
"asset": prediction.ticker,
"analysis_date": prediction.analysis_date.isoformat(),
"initial_price": prediction.initial_price,
"timeframes": prediction_payload["timeframes"],
"prediction": prediction_payload["prediction"],
"actual_outcome": {outcome["timeframe"]: outcome for outcome in outcomes},
"score": {
"data_quality_score": prediction.data_quality_score,
"confidence": prediction.confidence,
"market_regime": prediction.market_regime,
"volatility_regime": prediction.volatility_regime,
},
"mistakes": mistakes,
"lessons_learned": [item["lesson"] for item in memory_updates],
"memory_updates": memory_updates,
"feedback_loop_audit": feedback_audit,
}
return json_safe(report)
def daily_guard(self, db: Session, requested_batch: int) -> dict:
start = datetime.combine(datetime.utcnow().date(), time.min)
today_predictions = int(db.scalar(select(func.coalesce(func.sum(LearningRun.predictions_created), 0)).where(LearningRun.started_at >= start)) or 0)
max_daily = max(0, int(settings.learning_max_daily_runs))
remaining = max(0, max_daily - today_predictions)
effective_batch = min(max(1, requested_batch), remaining) if remaining else 0
projected = today_predictions + effective_batch
allowed = effective_batch > 0
partial = allowed and effective_batch < requested_batch
return {
"allowed": allowed,
"today_predictions": today_predictions,
"requested_batch": requested_batch,
"effective_batch": effective_batch,
"remaining_daily_budget": remaining,
"projected_predictions": projected,
"max_daily_runs": max_daily,
"partial_batch": partial,
"reason": (
"within daily limit"
if allowed and not partial
else "partial batch: using remaining daily learning budget to keep training without overfitting"
if partial
else "daily learning budget exhausted; waiting for next UTC window instead of oversampling the same day"
),
}
class HistoricalSamplerService:
def __init__(self) -> None:
self.rng = random.Random(self.seed())
def blended_sample(self, db: Session, seen_samples: set[tuple[str, str]] | None = None, trigger: str = "manual") -> dict | None:
if trigger == "alpha_loss_replay":
return self.alpha_loss_sample(db, seen_samples=seen_samples) or self.random_sample(db, seen_samples=seen_samples)
roll = self.rng.random()
random_ratio = clamp_ratio(settings.learning_random_sample_ratio)
alpha_ratio = clamp_ratio(settings.learning_alpha_loss_sample_ratio)
factor_ratio = clamp_ratio(settings.learning_factor_focus_sample_ratio)
preservation_ratio = clamp_ratio(settings.learning_capital_preservation_sample_ratio)
total = max(0.01, random_ratio + alpha_ratio + factor_ratio + preservation_ratio)
random_cut = random_ratio / total
alpha_cut = random_cut + alpha_ratio / total
factor_cut = alpha_cut + factor_ratio / total
if roll < random_cut:
return self.random_sample(db, seen_samples=seen_samples)
if roll < alpha_cut:
return self.alpha_loss_sample(db, seen_samples=seen_samples) or self.random_sample(db, seen_samples=seen_samples)
if roll < factor_cut:
return self.focus_priority_sample(db, seen_samples=seen_samples) or self.random_sample(db, seen_samples=seen_samples)
return self.capital_preservation_sample(db, seen_samples=seen_samples) or self.random_sample(db, seen_samples=seen_samples)
def alpha_loss_sample(self, db: Session, seen_samples: set[tuple[str, str]] | None = None) -> dict | None:
missed = db.scalars(
select(MissedWinner)
.order_by(desc(MissedWinner.benchmark_relative_return), desc(MissedWinner.created_at))
.limit(120)
).all()
for row in missed:
asset = db.scalar(select(Asset).where(Asset.ticker == row.ticker, Asset.is_active.is_(True)).limit(1))
if not asset:
continue
sample = self.sample_for_asset(db, asset, preferred_date=as_date(row.decision_date), seen_samples=seen_samples)
if sample:
sample["sampling_reason"] = "alpha_loss_replay"
sample["missed_winner_id"] = row.id
sample["benchmark_relative_return"] = row.benchmark_relative_return
return sample
return None
def focus_priority_sample(self, db: Session, seen_samples: set[tuple[str, str]] | None = None) -> dict | None:
priorities = db.scalars(
select(LearningFocusPriority)
.where(LearningFocusPriority.status.in_(["proposed", "active"]))
.order_by(desc(LearningFocusPriority.expected_learning_value), desc(LearningFocusPriority.created_at))
.limit(80)
).all()
for priority in priorities:
target = str(priority.target or "").upper()
asset = db.scalar(select(Asset).where(Asset.ticker == target, Asset.is_active.is_(True)).limit(1))
if not asset:
asset = db.scalar(select(Asset).where(Asset.sector.ilike(priority.target), Asset.is_active.is_(True)).limit(1))
if not asset:
linked_trade = db.scalar(
select(TradingGameTrade)
.where(TradingGameTrade.setup_type == priority.target)
.order_by(desc(TradingGameTrade.created_at))
.limit(1)
)
asset = db.scalar(select(Asset).where(Asset.ticker == linked_trade.ticker, Asset.is_active.is_(True)).limit(1)) if linked_trade else None
if not asset:
continue
sample = self.sample_for_asset(db, asset, preferred_date=None, seen_samples=seen_samples)
if sample:
sample["sampling_reason"] = "learning_focus_priority"
sample["learning_focus_priority_id"] = priority.id
sample["priority_type"] = priority.priority_type
return sample
return None
def capital_preservation_sample(self, db: Session, seen_samples: set[tuple[str, str]] | None = None) -> dict | None:
rows = db.scalars(
select(CapitalPreservationAlpha)
.where(CapitalPreservationAlpha.missed_gain > CapitalPreservationAlpha.avoided_loss)
.order_by(desc(CapitalPreservationAlpha.missed_gain), desc(CapitalPreservationAlpha.created_at))
.limit(80)
).all()
for row in rows:
asset = db.scalar(select(Asset).where(Asset.ticker == row.ticker, Asset.is_active.is_(True)).limit(1))
if not asset:
continue
sample = self.sample_for_asset(db, asset, preferred_date=as_date(row.decision_date), seen_samples=seen_samples)
if sample:
sample["sampling_reason"] = "capital_preservation_replay"
sample["capital_preservation_alpha_id"] = row.id
return sample
return None
def random_sample(self, db: Session, seen_samples: set[tuple[str, str]] | None = None) -> dict | None:
universe = {item.strip().lower() for item in settings.learning_asset_universe.split(",") if item.strip()}
asset_types = []
if "stocks" in universe or "stock" in universe:
asset_types.append("Stock")
if "etfs" in universe or "etf" in universe:
asset_types.append("ETF")
if not asset_types:
asset_types = ["Stock", "ETF"]
candidates = db.scalars(select(Asset).where(Asset.is_active.is_(True), Asset.asset_type.in_(asset_types))).all()
self.rng.shuffle(candidates)
for asset in candidates:
sample = self.sample_for_asset(db, asset, preferred_date=None, seen_samples=seen_samples)
if sample:
return sample
return None
def sample_for_asset(self, db: Session, asset: Asset, preferred_date: date | None, seen_samples: set[tuple[str, str]] | None = None) -> dict | None:
min_rows = max(252, settings.learning_min_history_years * 252)
stats = db.execute(
select(func.count(PriceHistory.id), func.min(PriceHistory.date), func.max(PriceHistory.date)).where(PriceHistory.asset_id == asset.id)
).one()
count, first_date, last_date = int(stats[0] or 0), as_date(stats[1]), as_date(stats[2])
if count < min_rows or not first_date or not last_date:
return None
earliest = first_date + timedelta(days=max(365, settings.learning_min_history_years * 365))
latest = last_date - timedelta(days=TIMEFRAMES["long"]["horizon_days"] * 2 + 30)
if earliest >= latest:
return None
preferred_candidates: list[date] = []
if preferred_date and earliest <= preferred_date <= latest:
preferred_candidates.extend([preferred_date, preferred_date - timedelta(days=3), preferred_date + timedelta(days=3)])
span = (latest - earliest).days
preferred_candidates.extend(earliest + timedelta(days=self.rng.randint(0, max(1, span))) for _ in range(8))
for candidate_date in preferred_candidates:
analysis_date = nearest_trading_date(db, asset, max(earliest, min(candidate_date, latest)), latest)
if not analysis_date:
continue
key = (asset.ticker, analysis_date.isoformat())
if seen_samples and key in seen_samples:
continue
return {"asset": asset, "analysis_date": analysis_date, "first_price_date": first_date, "last_price_date": last_date, "sample_rows": count}
return None
def seed(self) -> int:
raw = settings.learning_random_seed or datetime.utcnow().strftime("%Y%m%d%H")
return int(hashlib.sha256(raw.encode("utf-8")).hexdigest()[:8], 16)
class PointInTimeDataService:
def context_for(self, db: Session, asset: Asset, analysis_date: date) -> dict:
past_rows = db.scalars(
select(PriceHistory)
.where(PriceHistory.asset_id == asset.id, PriceHistory.date <= analysis_date)
.order_by(PriceHistory.date)
).all()
future_rows = db.scalars(
select(PriceHistory)
.where(
PriceHistory.asset_id == asset.id,
PriceHistory.date > analysis_date,
PriceHistory.date <= analysis_date + timedelta(days=TIMEFRAMES["long"]["horizon_days"] * 2 + 30),
)
.order_by(PriceHistory.date)
).all()
past_frame = price_frame(past_rows)
future_frame = price_frame(future_rows)
technical = TechnicalAnalysisEngine().analyze(past_frame, timeframe="1Y")
news = self.news_as_of(db, asset, analysis_date)
fundamentals = self.fundamentals_as_of(db, asset, analysis_date)
macro = self.macro_as_of(db, analysis_date)
market_context = self.market_context(db, analysis_date, past_frame)
initial_price = float(past_frame["close"].iloc[-1]) if not past_frame.empty else None
data_quality = data_quality_score(past_frame, news, fundamentals, macro)
return {
"asset": serialize_asset(asset),
"analysis_date": analysis_date.isoformat(),
"initial_price": round(initial_price, 4) if initial_price else None,
"past_prices": past_frame,
"future_prices": future_frame,
"technical": technical,
"news": news,
"fundamentals": fundamentals,
"macro": macro,
"market_context": market_context,
"data_quality_score": data_quality,
"evaluation_data_availability": {"future_ohlcv_rows": len(future_frame)},
"point_in_time_policy": {
"prices": "Only stored OHLCV rows with date <= analysis_date are used for prediction.",
"news": "Only linked public news with published_at <= analysis_date are used for prediction.",
"fundamentals": "Fundamentals are used only when filing dates are point-in-time verified; otherwise they are marked unavailable.",
"future": "Rows after analysis_date are hidden from PredictionEngine and used only by OutcomeEvaluator.",
},
}
def news_as_of(self, db: Session, asset: Asset, analysis_date: date) -> dict:
cutoff = datetime.combine(analysis_date, time.max)
rows = db.execute(
select(NewsArticle)
.join(NewsAssetLink, NewsAssetLink.article_id == NewsArticle.id)
.where(NewsAssetLink.asset_id == asset.id, NewsArticle.published_at <= cutoff)
.order_by(desc(NewsArticle.published_at))
.limit(80)
).scalars().all()
recent = [row for row in rows if row.published_at and row.published_at >= cutoff - timedelta(days=14)]
themes: dict[str, int] = defaultdict(int)
quality = []
for row in recent:
for theme in (row.theme_tags or {}).get("themes", []):
themes[str(theme)] += 1
quality.append(float(row.quality_score or 0.0))
return {
"article_count_total_as_of": len(rows),
"article_count_14d": len(recent),
"average_quality_14d": round(mean(quality), 3) if quality else 0.0,
"themes_14d": sorted([{"theme": key, "count": value} for key, value in themes.items()], key=lambda item: item["count"], reverse=True)[:8],
"sample_headlines": [
{
"title": row.title,
"source": row.source,
"published_at": row.published_at.isoformat() if row.published_at else None,
"themes": (row.theme_tags or {}).get("themes", []),
}
for row in recent[:8]
],
}
def fundamentals_as_of(self, db: Session, asset: Asset, analysis_date: date) -> dict:
snapshots = db.scalars(
select(FundamentalSnapshot)
.where(FundamentalSnapshot.asset_id == asset.id, FundamentalSnapshot.period_end <= analysis_date)
.order_by(desc(FundamentalSnapshot.period_end), desc(FundamentalSnapshot.created_at))
.limit(6)
).all()
for snapshot in snapshots:
filed = latest_filing_date(snapshot.metrics)
if filed and filed <= analysis_date:
return {
"status": "ready",
"provider": snapshot.provider,
"period_end": snapshot.period_end.isoformat() if snapshot.period_end else None,
"latest_filed": filed.isoformat(),
"quality_score": snapshot.quality_score,
"metrics": compact_metrics(snapshot.metrics),
}
return {
"status": "not_point_in_time_verified",
"quality_score": 0,
"message": "No stored fundamental snapshot had filing dates known by the simulated analysis date.",
}
def macro_as_of(self, db: Session, analysis_date: date) -> dict:
rows = db.scalars(select(MacroSnapshot).where(MacroSnapshot.date <= analysis_date).order_by(desc(MacroSnapshot.date)).limit(30)).all()
return {
"status": "ready" if rows else "missing",
"latest": [
{"indicator": row.indicator, "date": row.date.isoformat() if row.date else None, "value": row.value, "provider": row.provider}
for row in rows[:12]
],
}
def market_context(self, db: Session, analysis_date: date, asset_frame: pd.DataFrame) -> dict:
benchmark = db.scalar(select(Asset).where(Asset.ticker == settings.default_benchmark).limit(1))
benchmark_rows = []
if benchmark:
benchmark_rows = db.scalars(
select(PriceHistory)
.where(PriceHistory.asset_id == benchmark.id, PriceHistory.date <= analysis_date)
.order_by(PriceHistory.date)
).all()
frame = price_frame(benchmark_rows) if benchmark_rows else asset_frame
regime = infer_market_regime(frame)
volatility = infer_volatility_regime(frame)
return {
"benchmark": settings.default_benchmark if benchmark_rows else "asset_proxy",
"market_regime": regime,
"volatility_regime": volatility,
"market_breadth": "not_available_point_in_time",
"risk_sentiment": "risk_on" if regime in {"Bull Expansion", "Recovery", "Rotation"} else "risk_off" if regime in {"Risk-Off", "Panic"} else "balanced",
}
class PredictionEngine:
def predict(self, context: dict, db: Session | None = None, sample_metadata: dict | None = None) -> dict:
technical = context["technical"]
signal_scores = self.signal_scores(context)
sample_metadata = sample_metadata or {}
feedback = self.feedback_context(db, context, signal_scores, sample_metadata)
weights_used = feedback["weights_used"]
aggregate_score = weighted_score(signal_scores, weights_used)
dominant_direction = direction_from_score(aggregate_score, technical)
base_confidence = confidence_from_evidence(aggregate_score, context["data_quality_score"], context["market_context"], technical)
confidence = round(clamp(base_confidence + feedback["confidence_adjustment"], 15, 88), 1)
feedback["base_confidence"] = base_confidence
feedback["final_confidence"] = confidence
timeframes = {
timeframe: self.timeframe_prediction(timeframe, config, context, signal_scores, aggregate_score, dominant_direction, confidence)
for timeframe, config in TIMEFRAMES.items()
}
return {
"asset": context["asset"],
"analysis_date": context["analysis_date"],
"initial_price": context["initial_price"],
"prediction": {
"dominant_direction": dominant_direction,
"aggregate_confidence": confidence,
"aggregate_score": round(aggregate_score, 2),
"signal_scores": signal_scores,
"reasoning": self.reasoning(context, signal_scores, dominant_direction),
},
"timeframes": timeframes,
"feedback_loop": feedback,
"model_version_used": feedback["model_version_used"],
"weights_used": weights_used,
"learning_memory_used": feedback["learning_memory_used"],
"strategy_memory_used": feedback["strategy_memory_used"],
"research_priority_used": feedback["research_priority_used"],
"learning_mode_metadata": feedback["learning_mode_metadata"],
"anti_leakage": context["point_in_time_policy"],
}
def feedback_context(self, db: Session | None, context: dict, signal_scores: dict, sample_metadata: dict) -> dict:
model_version_used, weights_used, weight_source = active_weight_context(db)
signal_memory = signal_performance_context(db, context, signal_scores)
strategy_memory = strategy_memory_context(db, context)
research_priority = research_priority_context(db, sample_metadata)
mode_metadata = sample_metadata.get("learning_mode_metadata") or learning_mode_metadata(sample_metadata.get("run_trigger"), sample_metadata)
confidence_adjustment = round(
clamp(
signal_memory["confidence_delta"] + strategy_memory["confidence_delta"] + research_priority.get("confidence_delta", 0.0),
-14.0,
14.0,
),
2,
)
return {
"model_version_used": model_version_used,
"weight_source": weight_source,
"weights_used": weights_used,
"learning_memory_used": {
"signal_performance": signal_memory["rows"],
"confidence_delta": signal_memory["confidence_delta"],
"policy": "SignalPerformance reliability changes confidence only when enough outcome evidence exists.",
},
"strategy_memory_used": {
"rows": strategy_memory["rows"],
"confidence_delta": strategy_memory["confidence_delta"],
"policy": "StrategyMemory lessons modify confidence when their stored conditions match the current point-in-time setup.",
},
"research_priority_used": research_priority,
"learning_mode_metadata": mode_metadata,
"confidence_adjustment": confidence_adjustment,
"policy": "PredictionEngine uses active learned weights when available; otherwise BASE_SIGNAL_WEIGHTS. Learning memory changes confidence, not source code.",
}
def baseline_prediction(self, context: dict) -> dict:
signal_scores = self.signal_scores(context)
weights = normalize_weights(BASE_SIGNAL_WEIGHTS)
aggregate_score = weighted_score(signal_scores, weights)
direction = direction_from_score(aggregate_score, context["technical"])
confidence = confidence_from_evidence(aggregate_score, context["data_quality_score"], context["market_context"], context["technical"])
return {
"model_version_used": "base-static",
"weights_used": weights,
"aggregate_score": round(aggregate_score, 2),
"aggregate_confidence": confidence,
"dominant_direction": direction,
"actionability": feedback_actionability(direction, confidence, aggregate_score),
"confidence_adjustment": 0.0,
"memory_adjustment_used": False,
"policy": "Counterfactual baseline uses BASE_SIGNAL_WEIGHTS and ignores learned memory/confidence adjustments.",
}
def signal_scores(self, context: dict) -> dict:
technical = context["technical"]
indicators = technical.get("technical_indicators") or {}
volume = technical.get("volume") or {}
volatility = technical.get("volatility") or {}
levels = technical.get("levels") or {}
fundamentals = context["fundamentals"]
news = context["news"]
regime = context["market_context"]["market_regime"]
return {
"trend_structure": score_trend(technical),
"momentum": score_momentum(indicators),
"volume_confirmation": clamp(45 + safe_float(volume.get("relative_volume")) * 20),
"volatility_control": score_volatility(volatility),
"support_resistance": score_levels(levels),
"sentiment": clamp(45 + news.get("average_quality_14d", 0) * 18 + min(12, news.get("article_count_14d", 0) * 1.5)),
"narrative": clamp(40 + min(35, news.get("article_count_14d", 0) * 2.2) + min(20, len(news.get("themes_14d", [])) * 3)),
"fundamentals": safe_float(fundamentals.get("quality_score")) if fundamentals.get("status") == "ready" else 35.0,
"regime": 70.0 if regime in {"Bull Expansion", "Recovery", "Rotation"} else 42.0 if regime in {"Risk-Off", "Panic"} else 55.0,
}
def timeframe_prediction(self, timeframe: str, config: dict, context: dict, signal_scores: dict, aggregate: float, direction: str, confidence: float) -> dict:
price = safe_float(context["initial_price"])
technical = context["technical"]
levels = technical.get("levels") or {}
risk_reward = technical.get("risk_reward_estimate") or {}
horizon = config["horizon_days"]
scale = {"short": 0.65, "mid": 1.0, "long": 1.55}[timeframe]
base_move = max(1.5, abs(aggregate - 50) * 0.22 * scale)
if direction == "bullish":
move_range = [round(base_move, 2), round(base_move * 1.9, 2)]
target = price * (1 + move_range[1] / 100) if price else None
elif direction == "bearish":
move_range = [round(-base_move * 1.9, 2), round(-base_move, 2)]
target = price * (1 + move_range[0] / 100) if price else None
else:
move_range = [round(-base_move, 2), round(base_move, 2)]
target = levels.get("nearest_resistance") or levels.get("nearest_support")
invalidation = levels.get("invalidation_level") or (price * 0.94 if price else None)
if direction == "bearish" and price:
invalidation = levels.get("nearest_resistance") or price * 1.05
return {
"timeframe": timeframe,
"horizon_days": horizon,
"bias": direction,
"expected_move_percent": move_range,
"estimated_probability": round(confidence / 100, 3),
"entry_zone_informational": price_zone(price, levels, direction),
"invalidation_level": round_float(invalidation),
"target_zone": target_zone(target) if target else "not_available",
"risk": risk_label(context, signal_scores),
"confidence": round(confidence * (0.94 if timeframe == "long" else 1.0), 1),
"technical_reason": technical.get("technical_summary", "Technical evidence unavailable."),
"fundamental_reason": fundamental_reason(context["fundamentals"]),
"sentiment_reason": sentiment_reason(context["news"]),
"narrative_reason": narrative_reason(context["news"]),
"signals_used": sorted(signal_scores, key=signal_scores.get, reverse=True),
"risk_reward_estimate": risk_reward,
"validation_policy": "Forecast generated only from point-in-time context; future prices are hidden until outcome evaluation.",
}
def reasoning(self, context: dict, signal_scores: dict, direction: str) -> dict:
strongest = sorted(signal_scores.items(), key=lambda item: item[1], reverse=True)[:3]
weakest = sorted(signal_scores.items(), key=lambda item: item[1])[:3]
return {
"why_now": f"{context['asset']['ticker']} is simulated as {direction} because strongest point-in-time factors are {', '.join(key for key, _ in strongest)}.",
"supporting_evidence": [f"{key}: {value:.1f}/100" for key, value in strongest],
"contradicting_evidence": [f"{key}: {value:.1f}/100" for key, value in weakest if value < 50],
"missing_data": missing_data(context),
"intellectual_honesty": "The forecast is probabilistic. It is evaluated later against real future OHLCV and can be wrong.",
}
class OutcomeEvaluator:
def evaluate(self, context: dict, timeframe: str, prediction: dict) -> dict:
future = context["future_prices"].copy()
initial_price = safe_float(context["initial_price"])
horizon = int(prediction["horizon_days"])
analysis_date = parse_date(context["analysis_date"])
if future.empty or not initial_price:
return self.inconclusive(timeframe, horizon, "No future OHLCV rows are stored for the evaluation horizon.")
future["date"] = pd.to_datetime(future["date"]).dt.date
window = future.head(horizon).copy()
if window.empty:
return self.inconclusive(timeframe, horizon, "No future OHLCV rows exist before the target horizon.")
if len(window) < max(5, int(horizon * 0.70)):
return self.inconclusive(timeframe, horizon, "Future OHLCV rows are not mature enough for this horizon.")
evaluation_row = window.iloc[-1]
evaluation_price = safe_float(evaluation_row["close"])
realized = pct(initial_price, evaluation_price)
high_return = pct(initial_price, float(window["high"].astype(float).max()))
low_return = pct(initial_price, float(window["low"].astype(float).min()))
target_values = target_values_from_zone(prediction.get("target_zone"), initial_price)
invalidation = safe_float(prediction.get("invalidation_level"))
bias = prediction.get("bias", "neutral")
target_hit, time_to_target = hit_target(window, initial_price, bias, target_values)
invalidation_hit, time_to_invalidation = hit_invalidation(window, invalidation, bias)
direction_correct = classify_direction(realized, bias)
outcome_label = classify_prediction_outcome(direction_correct, target_hit, invalidation_hit, realized, bias)
confidence = safe_float(prediction.get("confidence")) / 100
realized_binary = 1.0 if direction_correct is True else 0.0 if direction_correct is False else 0.5
false_positive = bool(bias == "bullish" and outcome_label == "wrong")
false_negative = bool(bias in {"neutral", "bearish"} and high_return >= 8.0 and not target_hit)
missed = bool(false_negative or (bias == "neutral" and abs(realized) >= 8.0))
return {
"timeframe": timeframe,
"horizon_days": horizon,
"evaluation_date": evaluation_row["date"],
"price_at_evaluation": round(evaluation_price, 4),
"realized_return": round(realized, 3),
"max_favorable_excursion": round(high_return if bias != "bearish" else -low_return, 3),
"max_adverse_excursion": round(low_return if bias != "bearish" else -high_return, 3),
"drawdown": round(min(0.0, low_return), 3),
"target_hit": target_hit,
"invalidation_hit": invalidation_hit,
"time_to_target": time_to_target,
"time_to_invalidation": time_to_invalidation,
"direction_correct": direction_correct,
"false_positive": false_positive,
"false_negative": false_negative,
"missed_opportunity": missed,
"outcome_label": outcome_label,
"confidence_calibration_error": round(abs(confidence - realized_binary), 4),
"risk_reward_realized": realized_risk_reward(realized, low_return),
"why_right_or_wrong": why_outcome(outcome_label, bias, realized, target_hit, invalidation_hit),
}
def inconclusive(self, timeframe: str, horizon: int, reason: str) -> dict:
return {
"timeframe": timeframe,
"horizon_days": horizon,
"evaluation_date": None,
"price_at_evaluation": None,
"realized_return": None,
"max_favorable_excursion": None,
"max_adverse_excursion": None,
"drawdown": None,
"target_hit": False,
"invalidation_hit": False,
"direction_correct": None,
"false_positive": False,
"false_negative": False,
"missed_opportunity": False,
"outcome_label": "inconclusive",
"confidence_calibration_error": None,
"why_right_or_wrong": reason,
}
class MistakeAnalyzer:
def analyze(self, context: dict, prediction_payload: dict, prediction: dict, outcome: dict) -> dict:
outcome_label = outcome.get("outcome_label")
bias = prediction.get("bias")
confidence = safe_float(prediction.get("confidence"))
technical = context["technical"]
indicators = technical.get("technical_indicators") or {}
volume = technical.get("volume") or {}
volatility = technical.get("volatility") or {}
error_type = "random_market_noise"
severity = "Info"
if outcome_label == "inconclusive":
error_type = "insufficient_data"
elif outcome_label == "correct":
error_type = "underconfidence" if confidence < 50 else "random_market_noise"
elif outcome.get("invalidation_hit"):
error_type = "support_resistance_wrong"
severity = "Warning"
elif bias == "bullish" and safe_float(indicators.get("rsi")) >= 75:
error_type = "overbought_signal_ignored"
severity = "Warning"
elif bias == "bullish" and safe_float(volume.get("relative_volume")) < 1.1:
error_type = "weak_volume_confirmation"
severity = "Warning"
elif volatility.get("regime") == "high":
error_type = "volatility_expansion"
severity = "Warning"
elif confidence >= 70 and outcome_label == "wrong":
error_type = "overconfidence"
severity = "High"
elif context["news"].get("article_count_14d", 0) > 8 and outcome_label == "wrong":
error_type = "sentiment_overestimated"
severity = "Warning"
if error_type not in ERROR_TYPES and error_type not in {"correct_thesis_confirmed"}:
error_type = "random_market_noise"
explanation = mistake_explanation(error_type, context, prediction, outcome)
persist = outcome_label in {"wrong", "correct"} or error_type in {"overconfidence", "underconfidence", "insufficient_data"}
return {
"persist": persist,
"error_type": error_type,
"severity": severity,
"misleading_signal": misleading_signal(error_type),
"signal_to_weight_more": signal_to_weight_more(error_type),
"rule_adjustment": rule_adjustment(error_type),
"future_impact": future_impact(error_type),
"explanation": explanation,
"outcome_label": outcome_label,
}
class StrategyMemoryService:
def update_from_outcome(self, db: Session, context: dict, prediction: dict, outcome: dict, analysis: dict) -> list[dict]:
lessons = lessons_for(context, prediction, outcome, analysis)
updates = []
for lesson in lessons:
key = stable_key(lesson["category"], lesson["lesson"])
row = db.scalar(select(StrategyMemory).where(StrategyMemory.memory_key == key).limit(1))
if row is None:
row = StrategyMemory(memory_key=key, category=lesson["category"], lesson=lesson["lesson"], conditions=lesson["conditions"])
db.add(row)
row.sample_count = int(row.sample_count or 0) + 1
if outcome.get("outcome_label") == "correct":
row.positive_count = int(row.positive_count or 0) + 1
elif outcome.get("outcome_label") == "wrong":
row.negative_count = int(row.negative_count or 0) + 1
row.reliability_score = memory_reliability(int(row.positive_count or 0), int(row.negative_count or 0), int(row.sample_count or 0))
row.evidence = merge_evidence(row.evidence, context, prediction, outcome, analysis)
row.last_seen_at = datetime.utcnow()
row.updated_at = datetime.utcnow()
updates.append({"memory_key": row.memory_key, "category": row.category, "lesson": row.lesson, "reliability_score": row.reliability_score})
return updates
def update_signal_performance(self, db: Session, context: dict, prediction_payload: dict, outcomes: list[dict]) -> None:
signals = prediction_payload["prediction"].get("signal_scores", {})
for signal_name, score in signals.items():
for outcome in outcomes:
timeframe = outcome["timeframe"]
regime = context["market_context"]["market_regime"]
row = db.scalar(
select(SignalPerformance)
.where(SignalPerformance.signal_name == signal_name, SignalPerformance.timeframe == timeframe, SignalPerformance.market_regime == regime)
.limit(1)
)
if row is None:
row = SignalPerformance(signal_name=signal_name, timeframe=timeframe, market_regime=regime)
db.add(row)
row.sample_count = int(row.sample_count or 0) + 1
if outcome.get("direction_correct") is True:
row.correct_count = int(row.correct_count or 0) + 1
if outcome.get("false_positive"):
row.false_positive_count = int(row.false_positive_count or 0) + 1
if outcome.get("false_negative"):
row.false_negative_count = int(row.false_negative_count or 0) + 1
returns = (row.evidence or {}).get("returns", [])
drawdowns = (row.evidence or {}).get("drawdowns", [])
if outcome.get("realized_return") is not None:
returns = (returns + [outcome["realized_return"]])[-200:]
if outcome.get("drawdown") is not None:
drawdowns = (drawdowns + [outcome["drawdown"]])[-200:]
positives = [value for value in returns if value > 0]
negatives = [abs(value) for value in returns if value < 0]
correct_rate = int(row.correct_count or 0) / max(1, int(row.sample_count or 0))
false_penalty = int(row.false_positive_count or 0) / max(1, int(row.sample_count or 0)) * 18
row.average_return = round(mean(returns), 4) if returns else None
row.average_drawdown = round(mean(drawdowns), 4) if drawdowns else None
row.profit_factor = round(sum(positives) / max(0.01, sum(negatives)), 4) if returns else None
row.reliability_score = round(clamp(35 + correct_rate * 55 + min(10, max(-8, (row.average_return or 0))) - false_penalty), 2)
row.weight_adjustment = round((row.reliability_score - 50) / 500, 4)
row.evidence = {
"returns": returns,
"drawdowns": drawdowns,
"last_signal_score": score,
"last_outcome": outcome.get("outcome_label"),
"policy": "Reliability changes gradually and is penalized for false positives.",
}
row.updated_at = datetime.utcnow()
class ModelScoreService:
def recalculate(self, db: Session) -> dict:
rows = db.scalars(select(SignalPerformance).order_by(desc(SignalPerformance.updated_at)).limit(600)).all()
previous = active_model_version(db)
anti = self.anti_overfitting_report(db)
eligible_rows = [row for row in rows if int(row.sample_count or 0) >= MIN_MODEL_VERSION_SIGNAL_SAMPLE]
if (
int(anti.get("sample_count", 0) or 0) < MIN_MODEL_VERSION_OUTCOMES
or len(eligible_rows) < MIN_MODEL_VERSION_SIGNAL_ROWS
):
return {
"status": "insufficient_evidence",
"version": previous.version if previous else None,
"active_version": previous.version if previous else None,
"thresholds": {
"min_outcomes": MIN_MODEL_VERSION_OUTCOMES,
"min_signal_rows": MIN_MODEL_VERSION_SIGNAL_ROWS,
"min_signal_sample": MIN_MODEL_VERSION_SIGNAL_SAMPLE,
},
"evidence": {"outcomes": anti.get("sample_count", 0), "eligible_signal_rows": len(eligible_rows)},
"anti_overfitting": anti,
"policy": "No ModelVersion is created until enough outcome and signal reliability evidence exists.",
}
previous_weights = previous.weights if previous else BASE_SIGNAL_WEIGHTS
new_weights = dict(BASE_SIGNAL_WEIGHTS)
for signal_name in new_weights:
matching = [row for row in rows if row.signal_name == signal_name and row.sample_count >= MIN_MODEL_VERSION_SIGNAL_SAMPLE]
if not matching:
continue
avg_reliability = mean(row.reliability_score for row in matching)
new_weights[signal_name] = max(0.03, new_weights[signal_name] + (avg_reliability - 50) / 1000)
new_weights = normalize_weights(new_weights)
if previous and max(abs(safe_float(new_weights.get(key)) - safe_float((previous.weights or {}).get(key))) for key in BASE_SIGNAL_WEIGHTS) < 0.001:
return {
"status": "stable",
"version": previous.version,
"weights": previous.weights,
"anti_overfitting": anti,
"policy": "No new ModelVersion created because learned weights did not materially change.",
}
version = f"learning-{datetime.utcnow().strftime('%Y%m%d%H%M%S')}"
row = ModelVersion(
version=version,
model_name="BLUM Learning Loop",
weights=new_weights,
previous_weights=previous_weights,
training_window={"source": "historical_predictions", "evaluation_mode": settings.learning_evaluation_mode},
validation_metrics=LearningDashboardService().aggregate_metrics(db),
anti_overfitting_report=anti,
change_log="Adaptive signal reliability weights recalculated from point-in-time historical outcomes. Source code unchanged.",
is_active=True,
)
if previous:
previous.is_active = False
db.add(row)
db.add(
LearningMetric(
metric_name="model_weight_recalibration",
timeframe="all",
metric_value=anti.get("robustness_score"),
sample_count=int(anti.get("sample_count", 0)),
payload={"version": version, "weights": new_weights, "anti_overfitting": anti},
)
)
return {"status": "updated", "version": version, "weights": new_weights, "anti_overfitting": anti}
def anti_overfitting_report(self, db: Session) -> dict:
total = int(db.scalar(select(func.count(PredictionOutcome.id)).where(PredictionOutcome.outcome_label.in_(["correct", "wrong", "neutral"]))) or 0)
tickers = int(db.scalar(select(func.count(func.distinct(HistoricalPrediction.ticker)))) or 0)
sectors = int(db.scalar(select(func.count(func.distinct(HistoricalPrediction.sector)))) or 0)
regimes = int(db.scalar(select(func.count(func.distinct(HistoricalPrediction.market_regime)))) or 0)
outcomes = db.scalars(select(PredictionOutcome).where(PredictionOutcome.outcome_label.in_(["correct", "wrong", "neutral"])).limit(5000)).all()
correct = sum(1 for row in outcomes if row.outcome_label == "correct")
wrong = sum(1 for row in outcomes if row.outcome_label == "wrong")
usable = correct + wrong
hit_rate = correct / usable if usable else None
suspicious = bool(hit_rate is not None and hit_rate > 0.82 and usable < 250)
coverage_score = min(100, tickers * 2.0 + sectors * 6.0 + regimes * 5.0)
sample_score = min(100, total / 10)
penalty = 25 if suspicious else 0
robustness = clamp((coverage_score * 0.45) + (sample_score * 0.45) + (0 if suspicious else 10) - penalty)
return {
"sample_count": total,
"distinct_tickers": tickers,
"distinct_sectors": sectors,
"distinct_regimes": regimes,
"hit_rate": round(hit_rate, 4) if hit_rate is not None else None,
"temporal_validation": settings.learning_evaluation_mode,
"walk_forward_validation": settings.learning_evaluation_mode == "walk_forward",
"train_test_policy": "Historical samples are point-in-time; future OHLCV is only used after prediction persistence.",
"overfitting_warning": suspicious,
"robustness_score": round(robustness, 2),
"policy": "Penalize perfect-looking strategies, small samples, narrow ticker coverage and single-regime optimization.",
}
class FeedbackLoopAuditService:
def record(
self,
db: Session,
prediction: HistoricalPrediction,
prediction_payload: dict,
outcomes: list[dict],
memory_updates: list[dict],
) -> dict:
feedback = prediction_payload.get("feedback_loop", {})
usable = [row for row in outcomes if row.get("outcome_label") in {"correct", "wrong", "neutral"}]
correct = sum(1 for row in usable if row.get("outcome_label") == "correct")
wrong = sum(1 for row in usable if row.get("outcome_label") == "wrong")
returns = [safe_float(row.get("realized_return")) for row in usable if row.get("realized_return") is not None]
evidence_grade = "medium" if len(usable) >= 3 else "low" if usable else "insufficient"
learned = {
"memory_updates": memory_updates[:12],
"signal_performance_used": (feedback.get("learning_memory_used") or {}).get("signal_performance", []),
"strategy_memory_used": (feedback.get("strategy_memory_used") or {}).get("rows", []),
}
counterfactual = self.counterfactual_audit(prediction, prediction_payload, outcomes)
comparison = counterfactual.get("outcome_comparison", {})
changed_decision = bool(
(counterfactual.get("differences") or {}).get("direction_changed")
or (counterfactual.get("differences") or {}).get("actionability_changed")
or safe_float((counterfactual.get("differences") or {}).get("score_delta")) != 0.0
or safe_float((counterfactual.get("differences") or {}).get("confidence_delta")) != 0.0
)
improved = bool(comparison.get("improvement_detected"))
changes = {
"model_version_used": feedback.get("model_version_used"),
"weights_used": feedback.get("weights_used"),
"confidence_adjustment": feedback.get("confidence_adjustment"),
"base_confidence": feedback.get("base_confidence"),
"final_confidence": feedback.get("final_confidence"),
"research_priority_used": feedback.get("research_priority_used"),
"learning_mode_metadata": feedback.get("learning_mode_metadata"),
"counterfactual_audit": counterfactual,
}
decision = {
"prediction_id": prediction.id,
"ticker": prediction.ticker,
"analysis_date": iso(prediction.analysis_date),
"direction": prediction.expected_direction,
"confidence": prediction.confidence,
"aggregate_score": (prediction_payload.get("prediction") or {}).get("aggregate_score"),
"actionability": counterfactual.get("learned_prediction", {}).get("actionability"),
}
outcome_payload = {
"correct": correct,
"wrong": wrong,
"neutral": sum(1 for row in usable if row.get("outcome_label") == "neutral"),
"average_realized_return": round(mean(returns), 4) if returns else None,
"outcomes": {row.get("timeframe"): row.get("outcome_label") for row in outcomes},
"baseline_direction_correct": comparison.get("baseline_direction_correct"),
"learned_direction_correct": comparison.get("learned_direction_correct"),
"baseline_actionability": comparison.get("baseline_actionability"),
"learned_actionability": comparison.get("learned_actionability"),
"baseline_would_trade": comparison.get("baseline_would_trade"),
"learned_would_trade": comparison.get("learned_would_trade"),
"avoided_loss": comparison.get("avoided_loss"),
"missed_gain": comparison.get("missed_gain"),
"improvement_reason": comparison.get("improvement_reason"),
}
summary = (
f"{prediction.ticker} used {feedback.get('model_version_used', 'base-static')} with "
f"confidence adjustment {safe_float(feedback.get('confidence_adjustment')):.2f}; "
f"changed_decision={changed_decision}, improved={improved}. "
f"reason={comparison.get('improvement_reason', 'counterfactual_not_available')}."
)
row = FeedbackLoopAudit(
prediction_id=prediction.id,
ticker=prediction.ticker,
model_version_used=feedback.get("model_version_used") or "base-static",
learned_knowledge_json=json_safe(learned),
changes_applied_json=json_safe(changes),
future_decision_json=json_safe(decision),
outcome_json=json_safe(outcome_payload),
improvement_detected=improved,
evidence_grade=evidence_grade,
summary=summary,
)
db.add(row)
return {
"status": "recorded",
"prediction_id": prediction.id,
"model_version_used": row.model_version_used,
"what_was_learned": learned,
"what_changed": changes,
"future_decision_used_change": decision,
"counterfactual_audit": counterfactual,
"outcome": outcome_payload,
"improvement_detected": improved,
"evidence_grade": evidence_grade,
"summary": summary,
}
def counterfactual_audit(self, prediction: HistoricalPrediction, prediction_payload: dict, outcomes: list[dict]) -> dict:
context = dict(prediction.point_in_time_context or {})
context.pop("future_prices", None)
baseline = PredictionEngine().baseline_prediction(context) if context else {}
learned_prediction = prediction_payload.get("prediction") or {}
baseline_actionability = baseline.get("actionability")
learned_actionability = feedback_actionability(
learned_prediction.get("dominant_direction"),
learned_prediction.get("aggregate_confidence"),
learned_prediction.get("aggregate_score"),
)
learned = {
"model_version_used": (prediction_payload.get("feedback_loop") or {}).get("model_version_used") or "base-static",
"weights_used": prediction_payload.get("weights_used") or {},
"aggregate_score": learned_prediction.get("aggregate_score"),
"aggregate_confidence": learned_prediction.get("aggregate_confidence"),
"dominant_direction": learned_prediction.get("dominant_direction"),
"actionability": learned_actionability,
"confidence_adjustment": (prediction_payload.get("feedback_loop") or {}).get("confidence_adjustment", 0.0),
"memory_adjustment_used": bool((prediction_payload.get("feedback_loop") or {}).get("confidence_adjustment")),
}
baseline_correct = direction_correctness_summary(baseline.get("dominant_direction"), outcomes)
learned_correct = direction_correctness_summary(learned.get("dominant_direction"), outcomes)
baseline_would_trade = feedback_would_trade(baseline_actionability)
learned_would_trade = feedback_would_trade(learned_actionability)
returns = [safe_float(row.get("realized_return")) for row in outcomes if row.get("realized_return") is not None]
average_return = mean(returns) if returns else 0.0
avoided_loss = round(abs(average_return), 4) if baseline_would_trade and not learned_would_trade and average_return < 0 else 0.0
missed_gain = round(average_return, 4) if baseline_would_trade and not learned_would_trade and average_return > 0 else 0.0
improvement_detected, improvement_reason = counterfactual_improvement_reason(
baseline_correct,
learned_correct,
baseline_would_trade,
learned_would_trade,
average_return,
avoided_loss,
missed_gain,
)
comparison = {
"score_delta": round(safe_float(learned.get("aggregate_score")) - safe_float(baseline.get("aggregate_score")), 4),
"confidence_delta": round(safe_float(learned.get("aggregate_confidence")) - safe_float(baseline.get("aggregate_confidence")), 4),
"direction_changed": baseline.get("dominant_direction") != learned.get("dominant_direction"),
"actionability_changed": baseline_actionability != learned_actionability,
}
return {
"baseline_prediction": baseline,
"learned_prediction": learned,
"differences": comparison,
"outcome_comparison": {
"outcome_labels": {row.get("timeframe"): row.get("outcome_label") for row in outcomes},
"realized_returns": {row.get("timeframe"): row.get("realized_return") for row in outcomes},
"learned_direction": learned.get("dominant_direction"),
"baseline_direction": baseline.get("dominant_direction"),
"baseline_direction_correct": baseline_correct["direction_correct"],
"learned_direction_correct": learned_correct["direction_correct"],
"baseline_correct_count": baseline_correct["correct_count"],
"learned_correct_count": learned_correct["correct_count"],
"baseline_actionability": baseline_actionability,
"learned_actionability": learned_actionability,
"baseline_would_trade": baseline_would_trade,
"learned_would_trade": learned_would_trade,
"avoided_loss": avoided_loss,
"missed_gain": missed_gain,
"improvement_detected": improvement_detected,
"improvement_reason": improvement_reason,
"policy": "Outcome is observed after prediction persistence; baseline is recomputed only from point-in-time context.",
},
}
def report(self, db: Session, limit: int = 20) -> dict:
rows = db.scalars(select(FeedbackLoopAudit).order_by(desc(FeedbackLoopAudit.created_at)).limit(limit)).all()
return {
"status": "ready" if rows else "insufficient_evidence",
"rows": [serialize_feedback_audit(row) for row in rows],
"policy": "FeedbackLoopAudit is persisted by background learning runs and never computed by GET page render.",
}
class LearningDashboardService:
def dashboard(self, db: Session) -> dict:
latest_run = db.scalar(select(LearningRun).order_by(desc(LearningRun.started_at)).limit(1))
metrics = self.aggregate_metrics(db)
from app.services.trading_game import TradingGameSimulator
trading_game = TradingGameSimulator().status(db)
return {
"status": "active" if settings.enable_learning_loop else "passive",
"configuration": {
"batch_size": settings.learning_batch_size,
"max_daily_runs": settings.learning_max_daily_runs,
"min_history_years": settings.learning_min_history_years,
"asset_universe": settings.learning_asset_universe,
"evaluation_mode": settings.learning_evaluation_mode,
},
"latest_run": serialize_run(latest_run) if latest_run else None,
"metrics": metrics,
"signal_performance": self.signal_performance(db),
"strategy_memory": self.strategy_memory(db),
"mistakes": self.mistake_summary(db),
"model_versions": [serialize_model_version(row) for row in db.scalars(select(ModelVersion).order_by(desc(ModelVersion.created_at)).limit(8)).all()],
"feedback_loop_audit": FeedbackLoopAuditService().report(db, limit=8),
"trading_game": trading_game,
"policy": "BLUM Learning Loop optimizes calibration and robustness, not artificial 100% winrate.",
}
def aggregate_metrics(self, db: Session) -> dict:
outcomes = db.scalars(select(PredictionOutcome).where(PredictionOutcome.outcome_label.in_(["correct", "wrong", "neutral"]))).all()
by_timeframe = {}
for timeframe in TIMEFRAMES:
subset = [row for row in outcomes if row.timeframe == timeframe]
by_timeframe[timeframe] = metric_block(subset)
all_metrics = metric_block(outcomes)
return {
"simulations": int(db.scalar(select(func.count(HistoricalPrediction.id))) or 0),
"outcomes": len(outcomes),
"accuracy": {key: value["accuracy"] for key, value in by_timeframe.items()},
"by_timeframe": by_timeframe,
"overall": all_metrics,
"confidence_calibration": confidence_calibration(outcomes),
}
def signal_performance(self, db: Session, limit: int = 16) -> list[dict]:
return [serialize_signal_performance(row) for row in db.scalars(select(SignalPerformance).order_by(desc(SignalPerformance.reliability_score)).limit(limit)).all()]
def strategy_memory(self, db: Session, limit: int = 16) -> list[dict]:
return [serialize_strategy_memory(row) for row in db.scalars(select(StrategyMemory).order_by(desc(StrategyMemory.reliability_score), desc(StrategyMemory.updated_at)).limit(limit)).all()]
def mistake_summary(self, db: Session) -> list[dict]:
rows = db.execute(select(MistakeAnalysis.error_type, func.count(MistakeAnalysis.id)).group_by(MistakeAnalysis.error_type).order_by(desc(func.count(MistakeAnalysis.id))).limit(16)).all()
return [{"error_type": row[0], "count": int(row[1])} for row in rows]
def runs(self, db: Session, limit: int = 50) -> list[dict]:
return [serialize_run(row) for row in db.scalars(select(LearningRun).order_by(desc(LearningRun.started_at)).limit(limit)).all()]
def predictions(self, db: Session, ticker: str | None = None, limit: int = 80) -> list[dict]:
query = select(HistoricalPrediction).order_by(desc(HistoricalPrediction.created_at)).limit(limit)
if ticker:
query = select(HistoricalPrediction).where(HistoricalPrediction.ticker == ticker.upper()).order_by(desc(HistoricalPrediction.created_at)).limit(limit)
return [serialize_prediction(row) for row in db.scalars(query).all()]
def chat_memory(self, db: Session, query: str, assets: list[Asset] | None = None, limit: int = 8) -> dict:
tickers = {asset.ticker for asset in assets or []}
terms = {token.lower() for token in query.split() if len(token) > 3}
memory_rows = db.scalars(select(StrategyMemory).order_by(desc(StrategyMemory.updated_at)).limit(200)).all()
ranked = []
for row in memory_rows:
text = f"{row.category} {row.lesson} {row.conditions}".lower()
score = sum(1 for token in terms if token in text)
if score or not terms:
ranked.append((score, row))
ranked.sort(key=lambda item: (item[0], item[1].reliability_score, item[1].updated_at), reverse=True)
perf_query = select(SignalPerformance).order_by(desc(SignalPerformance.reliability_score)).limit(limit)
return {
"strategy_memory": [serialize_strategy_memory(row) for _, row in ranked[:limit]],
"signal_reliability": [serialize_signal_performance(row) for row in db.scalars(perf_query).all()],
"ticker_recent_predictions": self.predictions(db, next(iter(tickers)) if tickers else None, limit=6) if tickers else [],
"summary": "Learning Loop memory is based on point-in-time historical simulations and should adjust confidence, not create certainty.",
}
def nearest_trading_date(db: Session, asset: Asset, target: date, latest: date) -> date | None:
row = db.scalar(
select(PriceHistory.date)
.where(PriceHistory.asset_id == asset.id, PriceHistory.date >= target, PriceHistory.date <= latest)
.order_by(PriceHistory.date)
.limit(1)
)
return as_date(row)
def price_frame(rows: list[PriceHistory]) -> pd.DataFrame:
if not rows:
return pd.DataFrame()
return pd.DataFrame(
[
{
"date": row.date,
"open": row.open if row.open is not None else row.close,
"high": row.high if row.high is not None else row.close,
"low": row.low if row.low is not None else row.close,
"close": row.close,
"volume": row.volume or 0,
}
for row in rows
]
).sort_values("date")
def latest_filing_date(metrics: dict) -> date | None:
dates = []
for value in (metrics or {}).values():
if isinstance(value, dict) and value.get("filed"):
parsed = parse_date(value.get("filed"))
if parsed:
dates.append(parsed)
return max(dates) if dates else None
def compact_metrics(metrics: dict) -> dict:
output = {}
for key in ["revenue", "net_income", "operating_income", "assets", "liabilities", "operating_cash_flow", "eps_diluted"]:
if key in metrics:
output[key] = metrics[key]
return output
def infer_market_regime(frame: pd.DataFrame) -> str:
if frame is None or frame.empty or len(frame) < 80:
return "Sideways"
close = frame["close"].astype(float)
ret_3m = pct(float(close.iloc[-64]), float(close.iloc[-1])) if len(close) > 64 else 0
ret_1m = pct(float(close.iloc[-22]), float(close.iloc[-1])) if len(close) > 22 else 0
vol = close.pct_change().tail(30).std() * math.sqrt(252) * 100
drawdown = (float(close.iloc[-1]) / float(close.tail(126).max()) - 1) * 100
if drawdown < -18 and vol > 35:
return "Panic"
if ret_3m < -8 or drawdown < -12:
return "Risk-Off"
if ret_3m > 12 and ret_1m > 2:
return "Bull Expansion"
if ret_3m > 6 and ret_1m < 0:
return "Bull Maturity"
if ret_3m > 0 and ret_1m > 4:
return "Rotation"
if ret_3m > 3:
return "Recovery"
return "Sideways"
def infer_volatility_regime(frame: pd.DataFrame) -> str:
if frame is None or frame.empty or len(frame) < 30:
return "Unknown"
vol = frame["close"].astype(float).pct_change().tail(30).std() * math.sqrt(252) * 100
if vol >= 45:
return "High"
if vol >= 25:
return "Medium"
return "Low"
def data_quality_score(past: pd.DataFrame, news: dict, fundamentals: dict, macro: dict) -> float:
score = min(40, len(past) / 12)
score += min(20, len(past.tail(252)) / 13)
score += min(15, news.get("article_count_total_as_of", 0) * 1.5)
score += 15 if fundamentals.get("status") == "ready" else 4
score += 10 if macro.get("status") == "ready" else 3
return round(clamp(score), 2)
def score_trend(technical: dict) -> float:
if technical.get("status") != "ready":
return 35.0
direction = technical.get("trend_direction")
alignment = (technical.get("moving_averages") or {}).get("alignment")
score = 50.0
if direction == "uptrend":
score += 22
elif direction in {"downtrend", "bearish"}:
score -= 18
if alignment == "bullish_stack":
score += 16
elif alignment == "bearish_stack":
score -= 15
return clamp(score)
def score_momentum(indicators: dict) -> float:
rsi = safe_float(indicators.get("rsi"))
macd = safe_float(indicators.get("macd_hist"))
score = 50 + max(-18, min(18, macd * 1.2))
if 48 <= rsi <= 68:
score += 12
elif 68 < rsi <= 76:
score += 4
elif rsi > 76:
score -= 10
elif rsi < 38:
score -= 8
return clamp(score)
def score_volatility(volatility: dict) -> float:
regime = volatility.get("regime")
if regime == "low":
return 68
if regime == "medium":
return 58
if regime == "high":
return 38
return 50
def score_levels(levels: dict) -> float:
support = safe_float(levels.get("nearest_support"))
resistance = safe_float(levels.get("nearest_resistance"))
if not support or not resistance:
return 45
spread = abs(resistance - support) / max(1, support) * 100
return clamp(48 + min(22, spread * 1.2))
def weighted_score(scores: dict, weights: dict) -> float:
total = sum(weights.values()) or 1
return sum(safe_float(scores.get(key)) * weight for key, weight in weights.items()) / total
def direction_from_score(score: float, technical: dict) -> str:
trend = technical.get("trend_direction")
if score >= 61 and trend != "downtrend":
return "bullish"
if score <= 42 or trend == "downtrend":
return "bearish"
return "neutral"
def confidence_from_evidence(score: float, data_quality: float, market_context: dict, technical: dict) -> float:
distance = abs(score - 50)
confidence = 38 + distance * 0.72 + data_quality * 0.25
if market_context.get("market_regime") in {"Panic", "Risk-Off"}:
confidence -= 6
if technical.get("status") != "ready":
confidence -= 18
return round(clamp(confidence, 15, 82), 1)
def price_zone(price: float | None, levels: dict, direction: str) -> str:
if not price:
return "not_available"
if direction == "bullish":
anchor = safe_float(levels.get("breakout_level")) or price
return f"{anchor * 0.99:.2f}-{anchor * 1.01:.2f}"
if direction == "bearish":
anchor = safe_float(levels.get("breakdown_level")) or price
return f"{anchor * 0.99:.2f}-{anchor * 1.01:.2f}"
return f"{price * 0.98:.2f}-{price * 1.02:.2f}"
def target_zone(target: float | None) -> str:
if not target:
return "not_available"
return f"{target * 0.99:.2f}-{target * 1.01:.2f}"
def target_values_from_zone(zone: str | None, initial_price: float) -> list[float]:
if not zone or zone == "not_available":
return []
values = []
for chunk in str(zone).replace("–", "-").split("-"):
try:
values.append(float(chunk.strip()))
except ValueError:
pass
return values or [initial_price]
def hit_target(window: pd.DataFrame, initial_price: float, bias: str, targets: list[float]) -> tuple[bool, int | None]:
if not targets:
return False, None
target = max(targets) if bias != "bearish" else min(targets)
for index, row in enumerate(window.itertuples(index=False), start=1):
if bias == "bearish" and safe_float(row.low) <= target:
return True, index
if bias != "bearish" and safe_float(row.high) >= target:
return True, index
return False, None
def hit_invalidation(window: pd.DataFrame, invalidation: float, bias: str) -> tuple[bool, int | None]:
if not invalidation:
return False, None
for index, row in enumerate(window.itertuples(index=False), start=1):
if bias == "bearish" and safe_float(row.high) >= invalidation:
return True, index
if bias != "bearish" and safe_float(row.low) <= invalidation:
return True, index
return False, None
def classify_direction(realized: float, bias: str) -> bool | None:
if bias == "bullish":
return realized > 1.0
if bias == "bearish":
return realized < -1.0
if abs(realized) <= 3.0:
return True
return None
def classify_prediction_outcome(direction_correct: bool | None, target_hit: bool, invalidation_hit: bool, realized: float, bias: str) -> str:
if invalidation_hit and direction_correct is False:
return "wrong"
if target_hit and direction_correct is not False:
return "correct"
if direction_correct is True:
return "correct"
if direction_correct is False:
return "wrong"
return "neutral"
def realized_risk_reward(realized: float, adverse: float) -> float | None:
risk = abs(min(adverse, 0))
return round(realized / risk, 4) if risk > 0 else None
def why_outcome(label: str, bias: str, realized: float, target_hit: bool, invalidation_hit: bool) -> str:
if label == "correct":
return f"The {bias} thesis was supported: realized return {realized:.2f}%, target_hit={target_hit}, invalidation_hit={invalidation_hit}."
if label == "wrong":
return f"The {bias} thesis failed: realized return {realized:.2f}%, target_hit={target_hit}, invalidation_hit={invalidation_hit}."
return f"The {bias} thesis produced an inconclusive/neutral outcome: realized return {realized:.2f}%."
def mistake_explanation(error_type: str, context: dict, prediction: dict, outcome: dict) -> str:
ticker = context["asset"]["ticker"]
if outcome.get("outcome_label") == "correct":
return f"BLUM was directionally correct on {ticker}; this strengthens similar evidence but does not prove a permanent edge."
return f"BLUM classified the {ticker} {prediction.get('timeframe')} setup as {prediction.get('bias')}, but outcome was {outcome.get('outcome_label')}. Primary error class: {error_type}."
def misleading_signal(error_type: str) -> str:
return {
"overbought_signal_ignored": "Momentum was treated as continuation even though RSI was stretched.",
"weak_volume_confirmation": "Price structure looked constructive but volume confirmation was weak.",
"sentiment_overestimated": "News/narrative intensity may have been over-weighted versus price confirmation.",
"support_resistance_wrong": "Invalidation/support-resistance zone was too close, stale or structurally weak.",
"volatility_expansion": "Volatility regime expanded faster than the setup could absorb.",
"overconfidence": "Confidence was too high relative to evidence diversity.",
"underconfidence": "Confidence was too low relative to realized follow-through.",
"insufficient_data": "Historical window lacked enough verified data for robust evaluation.",
}.get(error_type, "No single misleading signal can be isolated; random market noise remains possible.")
def signal_to_weight_more(error_type: str) -> str:
return {
"weak_volume_confirmation": "relative_volume and accumulation/distribution",
"sentiment_overestimated": "price/sentiment divergence and source quality",
"support_resistance_wrong": "ATR-adjusted invalidation and level recency",
"volatility_expansion": "ATR percent, realized volatility and regime instability",
"overconfidence": "contradicting evidence and sample-size penalty",
"underconfidence": "aligned trend/volume/fundamental confirmation",
}.get(error_type, "independent confirmations and out-of-sample sample size")
def rule_adjustment(error_type: str) -> str:
return {
"overbought_signal_ignored": "Reduce breakout confidence when RSI > 75 unless volume is expanding and pullback quality is acceptable.",
"weak_volume_confirmation": "Require relative volume > 1.3x for breakout confidence upgrades.",
"sentiment_overestimated": "Treat narrative/news as context until price and sector confirmation agree.",
"support_resistance_wrong": "Use ATR-adjusted buffers around invalidation rather than raw pivot levels only.",
"volatility_expansion": "Increase risk penalty in high-volatility regimes.",
"overconfidence": "Cap confidence when evidence comes from fewer than three independent factors.",
"underconfidence": "Allow moderate confidence upgrade when trend, volume and fundamentals are aligned out-of-sample.",
}.get(error_type, "Aggregate more similar cases before changing weights.")
def future_impact(error_type: str) -> str:
return f"Future BLUM scoring should adjust the relevant factor weight for {error_type} only after repeated out-of-sample confirmation."
def lessons_for(context: dict, prediction: dict, outcome: dict, analysis: dict) -> list[dict]:
technical = context["technical"]
indicators = technical.get("technical_indicators") or {}
volume = technical.get("volume") or {}
lessons = []
if safe_float(indicators.get("rsi")) > 75:
lessons.append(
{
"category": "momentum_risk",
"lesson": "RSI above 75 increases false-breakout risk unless volume and trend quality independently confirm.",
"conditions": {"rsi_gt": 75, "timeframe": prediction.get("timeframe")},
}
)
if safe_float(volume.get("relative_volume")) > 1.5:
lessons.append(
{
"category": "volume_confirmation",
"lesson": "Momentum breakout reliability improves when relative volume is above 1.5x.",
"conditions": {"relative_volume_gt": 1.5, "timeframe": prediction.get("timeframe")},
}
)
if context["fundamentals"].get("status") != "ready" and prediction.get("bias") == "bullish":
lessons.append(
{
"category": "fundamental_gap",
"lesson": "Bullish technical setups without point-in-time verified fundamentals need lower confidence.",
"conditions": {"fundamentals": "not_point_in_time_verified"},
}
)
if analysis.get("error_type") in {"sentiment_overestimated", "weak_volume_confirmation", "volatility_expansion"}:
lessons.append(
{
"category": analysis["error_type"],
"lesson": rule_adjustment(analysis["error_type"]),
"conditions": {"error_type": analysis["error_type"], "timeframe": prediction.get("timeframe")},
}
)
if not lessons:
lessons.append(
{
"category": "general_calibration",
"lesson": "Use repeated out-of-sample outcomes before changing confidence materially.",
"conditions": {"outcome_label": outcome.get("outcome_label"), "timeframe": prediction.get("timeframe")},
}
)
return lessons
def memory_reliability(positive: int, negative: int, sample_count: int) -> float:
usable = positive + negative
if usable == 0:
return 50.0
rate = positive / usable
sample_bonus = min(12, math.log1p(sample_count) * 4)
return round(clamp(35 + rate * 50 + sample_bonus - (negative / max(1, sample_count)) * 12), 2)
def merge_evidence(existing: dict, context: dict, prediction: dict, outcome: dict, analysis: dict) -> dict:
samples = (existing or {}).get("samples", [])
samples.append(
{
"ticker": context["asset"]["ticker"],
"analysis_date": context["analysis_date"],
"timeframe": prediction.get("timeframe"),
"bias": prediction.get("bias"),
"outcome": outcome.get("outcome_label"),
"realized_return": outcome.get("realized_return"),
"error_type": analysis.get("error_type"),
}
)
return {"samples": samples[-40:], "last_update_policy": "Evidence is append-only and capped to latest samples for compact storage."}
def active_model_version(db: Session) -> ModelVersion | None:
return db.scalar(select(ModelVersion).where(ModelVersion.is_active.is_(True)).order_by(desc(ModelVersion.created_at)).limit(1))
def active_weight_context(db: Session | None) -> tuple[str, dict, str]:
if db is not None:
row = active_model_version(db)
if row and isinstance(row.weights, dict) and row.weights:
return row.version, model_weights_with_fallback(row.weights), "active_model_version"
return "base-static", normalize_weights(BASE_SIGNAL_WEIGHTS), "base_signal_weights"
def learning_mode_metadata(trigger: str | None, sample_metadata: dict | None = None) -> dict:
sample_metadata = sample_metadata or {}
sampling_reason = sample_metadata.get("sampling_reason") or "random_point_in_time"
mode = "training_replay" if sampling_reason in {"alpha_loss_replay", "learning_focus_priority", "capital_preservation_replay"} or trigger == "alpha_loss_replay" else "walk_forward_validation"
if trigger == "paper_forward" or sample_metadata.get("mode") == "paper_forward":
mode = "paper_forward"
return {
"mode": mode,
"training_replay": mode == "training_replay",
"walk_forward_validation": mode == "walk_forward_validation",
"paper_forward": mode == "paper_forward",
"trigger": trigger or sample_metadata.get("run_trigger") or "unknown",
"sampling_reason": sampling_reason,
"evaluation_mode": sample_metadata.get("evaluation_mode") or settings.learning_evaluation_mode,
"policy": "Mode metadata is descriptive. It changes audit traceability, not source code or frontend execution.",
}
def feedback_actionability(direction: str | None, confidence: float | None, score: float | None = None) -> str:
confidence_value = safe_float(confidence)
score_value = safe_float(score)
if direction == "neutral" or confidence_value < 35:
return "watch"
if confidence_value >= 68 and direction in {"bullish", "bearish"} and abs(score_value - 50) >= 10:
return "active_setup"
if confidence_value >= 54 and direction in {"bullish", "bearish"}:
return "wait_for_trigger"
return "watch"
def feedback_would_trade(actionability: str | None) -> bool:
return actionability in {"active_setup", "wait_for_trigger", "actionable_if_confirmed"}
def direction_correctness_summary(direction: str | None, outcomes: list[dict]) -> dict:
rows = [row for row in outcomes if row.get("realized_return") is not None]
if not rows:
return {"direction_correct": None, "correct_count": 0, "wrong_count": 0, "sample_count": 0}
correct = sum(1 for row in rows if direction_matches_return(direction, safe_float(row.get("realized_return"))))
wrong = len(rows) - correct
return {
"direction_correct": correct > wrong,
"correct_count": correct,
"wrong_count": wrong,
"sample_count": len(rows),
}
def direction_matches_return(direction: str | None, realized_return: float) -> bool:
if direction == "bullish":
return realized_return > 0
if direction == "bearish":
return realized_return < 0
if direction == "neutral":
return abs(realized_return) <= 1.0
return False
def counterfactual_improvement_reason(
baseline_correct: dict,
learned_correct: dict,
baseline_would_trade: bool,
learned_would_trade: bool,
average_return: float,
avoided_loss: float,
missed_gain: float,
) -> tuple[bool, str]:
baseline_count = int(baseline_correct.get("correct_count") or 0)
learned_count = int(learned_correct.get("correct_count") or 0)
if learned_count > baseline_count:
return True, "learned_direction_more_correct_than_baseline"
if learned_count < baseline_count:
return False, "learned_direction_less_correct_than_baseline"
if avoided_loss > 0:
return True, "learned_actionability_avoided_baseline_loss"
if missed_gain > 0:
return False, "learned_actionability_missed_baseline_gain"
if learned_would_trade and not baseline_would_trade and average_return > 0:
return True, "learned_actionability_captured_gain_baseline_would_skip"
if learned_would_trade and not baseline_would_trade and average_return < 0:
return False, "learned_actionability_entered_losing_trade_baseline_would_skip"
return False, "no_counterfactual_improvement_detected"
def model_weights_with_fallback(weights: dict) -> dict:
merged = dict(BASE_SIGNAL_WEIGHTS)
for key in BASE_SIGNAL_WEIGHTS:
if key in weights:
merged[key] = safe_float(weights.get(key)) if weights.get(key) is not None else BASE_SIGNAL_WEIGHTS[key]
return normalize_weights(merged)
def signal_performance_context(db: Session | None, context: dict, signal_scores: dict) -> dict:
if db is None or not signal_scores:
return {"rows": [], "confidence_delta": 0.0}
regime = context.get("market_context", {}).get("market_regime", "Unknown")
rows = db.scalars(
select(SignalPerformance)
.where(SignalPerformance.signal_name.in_(list(signal_scores.keys())), SignalPerformance.market_regime == regime)
.order_by(desc(SignalPerformance.sample_count), desc(SignalPerformance.updated_at))
.limit(80)
).all()
grouped: dict[str, list[SignalPerformance]] = defaultdict(list)
for row in rows:
grouped[row.signal_name].append(row)
payloads = []
deltas = []
for signal_name, signal_rows in grouped.items():
sample_count = sum(int(row.sample_count or 0) for row in signal_rows)
reliability = mean(float(row.reliability_score or 50.0) for row in signal_rows)
false_positives = sum(int(row.false_positive_count or 0) for row in signal_rows)
false_positive_rate = false_positives / max(1, sample_count)
enough_evidence = sample_count >= MIN_MODEL_VERSION_SIGNAL_SAMPLE
delta = clamp((reliability - 50.0) / 8.0 - false_positive_rate * 4.0, -5.0, 5.0) if enough_evidence else 0.0
payloads.append(
{
"signal_name": signal_name,
"market_regime": regime,
"sample_count": sample_count,
"reliability_score": round(reliability, 2),
"false_positive_rate": round(false_positive_rate, 4),
"signal_score": signal_scores.get(signal_name),
"confidence_delta": round(delta, 3),
"used": enough_evidence,
}
)
if enough_evidence:
deltas.append(delta)
total_delta = clamp(sum(deltas) / max(1, len(deltas)) * 1.4, -8.0, 8.0) if deltas else 0.0
return {"rows": payloads[:16], "confidence_delta": round(total_delta, 3)}
def strategy_memory_context(db: Session | None, context: dict) -> dict:
if db is None:
return {"rows": [], "confidence_delta": 0.0}
rows = db.scalars(select(StrategyMemory).order_by(desc(StrategyMemory.reliability_score), desc(StrategyMemory.updated_at)).limit(120)).all()
applicable = []
deltas = []
for row in rows:
if not strategy_memory_matches(row, context):
continue
sample_count = int(row.sample_count or 0)
enough_evidence = sample_count >= 3
positive = int(row.positive_count or 0)
negative = int(row.negative_count or 0)
reliability = safe_float(row.reliability_score) if row.reliability_score is not None else 50.0
delta = clamp((reliability - 50.0) / 10.0, -4.0, 4.0) if enough_evidence else 0.0
if negative > positive and enough_evidence:
delta = min(delta, -1.5)
applicable.append(
{
"memory_key": row.memory_key,
"category": row.category,
"lesson": row.lesson,
"sample_count": sample_count,
"positive_count": positive,
"negative_count": negative,
"reliability_score": reliability,
"confidence_delta": round(delta, 3),
"used": enough_evidence,
}
)
if enough_evidence:
deltas.append(delta)
total_delta = clamp(sum(deltas), -6.0, 6.0) if deltas else 0.0
return {"rows": applicable[:12], "confidence_delta": round(total_delta, 3)}
def strategy_memory_matches(row: StrategyMemory, context: dict) -> bool:
conditions = row.conditions or {}
technical = context.get("technical") or {}
indicators = technical.get("technical_indicators") or {}
volume = technical.get("volume") or {}
fundamentals = context.get("fundamentals") or {}
if "rsi_gt" in conditions and safe_float(indicators.get("rsi")) > safe_float(conditions.get("rsi_gt")):
return True
if "relative_volume_gt" in conditions and safe_float(volume.get("relative_volume")) > safe_float(conditions.get("relative_volume_gt")):
return True
if conditions.get("fundamentals") == "not_point_in_time_verified" and fundamentals.get("status") != "ready":
return True
if row.category in {"general_calibration", "volume_confirmation"}:
return True
return False
def research_priority_context(db: Session | None, sample_metadata: dict) -> dict:
priority_id = sample_metadata.get("learning_focus_priority_id")
if db is not None and priority_id:
row = db.get(LearningFocusPriority, priority_id)
if row:
return {
"status": "used",
"id": row.id,
"priority_type": row.priority_type,
"target": row.target,
"reason": row.reason,
"expected_learning_value": row.expected_learning_value,
"urgency": row.urgency,
"sample_gap": row.sample_gap,
"sampling_reason": sample_metadata.get("sampling_reason"),
"confidence_delta": 0.0,
}
if sample_metadata.get("sampling_reason"):
return {
"status": "used",
"sampling_reason": sample_metadata.get("sampling_reason"),
"priority_type": sample_metadata.get("priority_type"),
"missed_winner_id": sample_metadata.get("missed_winner_id"),
"confidence_delta": 0.0,
}
return {"status": "not_applicable", "confidence_delta": 0.0}
def normalize_weights(weights: dict[str, float]) -> dict[str, float]:
total = sum(max(0.0, float(value)) for value in weights.values()) or 1.0
return {key: round(max(0.0, float(value)) / total, 4) for key, value in weights.items()}
def metric_block(rows: list[PredictionOutcome]) -> dict:
if not rows:
return {"sample_count": 0, "accuracy": None, "average_return": None, "average_drawdown": None, "profit_factor": None, "sharpe_proxy": None, "max_drawdown": None}
correct = sum(1 for row in rows if row.outcome_label == "correct")
wrong = sum(1 for row in rows if row.outcome_label == "wrong")
usable = correct + wrong
returns = [float(row.realized_return) for row in rows if row.realized_return is not None]
drawdowns = [float(row.drawdown) for row in rows if row.drawdown is not None]
positives = [value for value in returns if value > 0]
negatives = [abs(value) for value in returns if value < 0]
return {
"sample_count": len(rows),
"accuracy": round(correct / usable, 4) if usable else None,
"win_rate": round(correct / max(1, len(rows)), 4),
"average_return": round(mean(returns), 4) if returns else None,
"average_drawdown": round(mean(drawdowns), 4) if drawdowns else None,
"profit_factor": round(sum(positives) / max(0.01, sum(negatives)), 4) if returns else None,
"sharpe_proxy": round(mean(returns) / max(0.01, stdev(returns)), 4) if len(returns) > 2 else None,
"max_drawdown": round(min(drawdowns), 4) if drawdowns else None,
"false_positive_rate": round(sum(1 for row in rows if row.false_positive) / max(1, len(rows)), 4),
}
def confidence_calibration(rows: list[PredictionOutcome]) -> dict:
values = [row.confidence_calibration_error for row in rows if row.confidence_calibration_error is not None]
if not values:
return {"status": "insufficient_sample", "mean_absolute_error": None, "sample_count": 0}
error = mean(float(value) for value in values)
return {"status": "calibrated" if error <= 0.28 else "needs_attention", "mean_absolute_error": round(error, 4), "sample_count": len(values)}
def serialize_run(row: LearningRun | None) -> dict | None:
if row is None:
return None
return {
"run_id": row.run_id,
"trigger": row.trigger,
"status": row.status,
"evaluation_mode": row.evaluation_mode,
"batch_size": row.batch_size,
"predictions_created": row.predictions_created,
"outcomes_evaluated": row.outcomes_evaluated,
"mistakes_found": row.mistakes_found,
"memory_updates": row.memory_updates,
"started_at": iso(row.started_at),
"completed_at": iso(row.completed_at),
"summary": row.summary,
"anti_overfitting_report": row.anti_overfitting_report,
}
def serialize_prediction(row: HistoricalPrediction) -> dict:
return {
"id": row.id,
"ticker": row.ticker,
"asset_type": row.asset_type,
"sector": row.sector,
"market": row.market,
"analysis_date": iso(row.analysis_date),
"initial_price": row.initial_price,
"expected_direction": row.expected_direction,
"confidence": row.confidence,
"market_regime": row.market_regime,
"volatility_regime": row.volatility_regime,
"data_quality_score": row.data_quality_score,
"model_version_used": row.model_version_used,
"weights_used": row.weights_used,
"learning_memory_used": row.learning_memory_used,
"strategy_memory_used": row.strategy_memory_used,
"research_priority_used": row.research_priority_used,
"prediction": row.prediction_payload.get("prediction", {}) if row.prediction_payload else {},
"timeframes": row.prediction_payload.get("timeframes", {}) if row.prediction_payload else {},
"created_at": iso(row.created_at),
}
def serialize_signal_performance(row: SignalPerformance) -> dict:
return {
"signal_name": row.signal_name,
"timeframe": row.timeframe,
"market_regime": row.market_regime,
"sample_count": row.sample_count,
"correct_count": row.correct_count,
"false_positive_count": row.false_positive_count,
"false_negative_count": row.false_negative_count,
"average_return": row.average_return,
"average_drawdown": row.average_drawdown,
"profit_factor": row.profit_factor,
"reliability_score": row.reliability_score,
"weight_adjustment": row.weight_adjustment,
"updated_at": iso(row.updated_at),
}
def serialize_strategy_memory(row: StrategyMemory) -> dict:
return {
"memory_key": row.memory_key,
"category": row.category,
"lesson": row.lesson,
"conditions": row.conditions,
"reliability_score": row.reliability_score,
"sample_count": row.sample_count,
"positive_count": row.positive_count,
"negative_count": row.negative_count,
"last_seen_at": iso(row.last_seen_at),
}
def serialize_model_version(row: ModelVersion) -> dict:
return {
"version": row.version,
"model_name": row.model_name,
"is_active": row.is_active,
"weights": row.weights,
"validation_metrics": row.validation_metrics,
"anti_overfitting_report": row.anti_overfitting_report,
"change_log": row.change_log,
"created_at": iso(row.created_at),
}
def serialize_feedback_audit(row: FeedbackLoopAudit) -> dict:
return {
"id": row.id,
"prediction_id": row.prediction_id,
"ticker": row.ticker,
"model_version_used": row.model_version_used,
"what_was_learned": row.learned_knowledge_json,
"what_changed": row.changes_applied_json,
"counterfactual_audit": (row.changes_applied_json or {}).get("counterfactual_audit"),
"future_decision_used_change": row.future_decision_json,
"outcome": row.outcome_json,
"improvement_detected": row.improvement_detected,
"evidence_grade": row.evidence_grade,
"summary": row.summary,
"created_at": iso(row.created_at),
}
def fundamental_reason(fundamentals: dict) -> str:
if fundamentals.get("status") == "ready":
return f"Point-in-time verified fundamentals available with quality score {fundamentals.get('quality_score')}."
return "Fundamentals are not point-in-time verified for this simulated date and are not used as hard evidence."
def sentiment_reason(news: dict) -> str:
return f"{news.get('article_count_14d', 0)} linked headlines in the prior 14 days; average source quality {news.get('average_quality_14d', 0)}."
def narrative_reason(news: dict) -> str:
themes = ", ".join(item["theme"] for item in news.get("themes_14d", [])[:4]) or "no dominant stored theme"
return f"Dominant point-in-time themes: {themes}."
def risk_label(context: dict, scores: dict) -> str:
if context["market_context"]["market_regime"] in {"Panic", "Risk-Off"} or scores.get("volatility_control", 50) < 42:
return "High"
if min(scores.values()) < 42:
return "Medium"
return "Low/Medium"
def missing_data(context: dict) -> list[str]:
missing = []
if context["fundamentals"].get("status") != "ready":
missing.append("point-in-time verified fundamentals")
if context["news"].get("article_count_total_as_of", 0) == 0:
missing.append("linked historical news")
if context["macro"].get("status") != "ready":
missing.append("macro snapshots")
return missing or ["no critical missing point-in-time field"]
def serialize_asset(asset: Asset) -> dict:
return {
"ticker": asset.ticker,
"name": asset.name,
"asset_type": asset.asset_type,
"sector": asset.sector,
"industry": asset.industry,
"country": asset.country,
"currency": asset.currency,
"exchange": asset.exchange,
}
def stable_key(category: str, lesson: str) -> str:
return f"{category}:{hashlib.sha256(lesson.encode('utf-8')).hexdigest()[:18]}"
def parse_date(value) -> date | None:
if value is None:
return None
if isinstance(value, datetime):
return value.date()
if isinstance(value, date):
return value
try:
return datetime.fromisoformat(str(value)[:10]).date()
except Exception:
return None
def as_date(value) -> date | None:
return parse_date(value)
def pct(start: float, end: float) -> float:
return ((end / start) - 1) * 100 if start else 0.0
def safe_float(value) -> float:
try:
if value is None:
return 0.0
numeric = float(value)
return numeric if math.isfinite(numeric) else 0.0
except (TypeError, ValueError):
return 0.0
def clamp(value: float, low: float = 0.0, high: float = 100.0) -> float:
return max(low, min(high, float(value)))
def clamp_ratio(value: float) -> float:
return max(0.0, min(1.0, safe_float(value)))
def round_float(value) -> float | None:
return round(safe_float(value), 4) if value is not None else None
def iso(value) -> str | None:
return value.isoformat() if hasattr(value, "isoformat") else str(value) if value is not None else None
def json_safe(value):
if value is None or isinstance(value, (str, bool, int)):
return value
if isinstance(value, float):
return value if math.isfinite(value) else None
if isinstance(value, (datetime, date)):
return value.isoformat()
if isinstance(value, dict):
return {str(key): json_safe(item) for key, item in value.items() if key not in {"past_prices", "future_prices"}}
if isinstance(value, (list, tuple, set)):
return [json_safe(item) for item in value]
if hasattr(value, "item"):
try:
return json_safe(value.item())
except Exception:
pass
return str(value)
|