Spaces:
Sleeping
Sleeping
File size: 45,253 Bytes
cb488aa 824e8e7 5117aff 824e8e7 cb488aa 824e8e7 5117aff 824e8e7 5117aff 824e8e7 cb488aa 0fb2300 cb488aa 5117aff cb488aa 5117aff cb488aa 5117aff cb488aa 5117aff cb488aa 824e8e7 cb488aa 824e8e7 5117aff 824e8e7 5117aff 824e8e7 5117aff 824e8e7 5117aff 824e8e7 5117aff 824e8e7 | 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 | import os
import json
import logging
import joblib
import uvicorn
import numpy as np
import pandas as pd
import yfinance as yf
import tensorflow as tf
from fastapi import FastAPI, HTTPException
from fastapi.middleware.cors import CORSMiddleware
from pydantic import BaseModel, Field
from typing import List, Dict, Any, Optional
from huggingface_hub import hf_hub_download
from datetime import datetime, timedelta
from dotenv import load_dotenv
from groq import Groq
# Load environment variables
load_dotenv()
# Configure logging
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)
# --- LLM Client Setup ---
llm_client = Groq(
api_key=os.getenv("GROQ_API_KEY")
)
def analyze_cascade_with_llm(
nodes: List[str],
shocked_node: str,
shock_magnitude: float,
connectivity_strength: float,
liquidity_buffer: float,
systemic_risk: float,
congestion_level: str
) -> Dict[str, Any]:
"""
Use LLM to analyze cascade with HARD BOUNDARIES.
Cascade triggers deterministically based on thresholds:
- Shock magnitude > 0.6 (60%)
- Connectivity strength > 1.3
- Liquidity buffer < 0.7
- Systemic risk > 0.5 (50%)
"""
# HARD BOUNDARY RULES - cascade triggers if ANY of these:
trigger_shock = shock_magnitude > 0.6
trigger_connectivity = connectivity_strength > 1.3
trigger_liquidity = liquidity_buffer < 0.7
trigger_systemic = systemic_risk > 0.5
trigger_congestion = congestion_level == "high"
should_cascade = (trigger_shock and (trigger_connectivity or trigger_liquidity)) or \
(trigger_systemic and trigger_congestion) or \
(shock_magnitude > 0.8) # Very high shock always cascades
if not should_cascade:
return {
"status": "stable",
"failed_nodes": [],
"failed_edges": [],
"failure_ratio": 0.0,
"cascade_depth": 0,
"analysis": "Network is stable. No cascade triggered."
}
# Use LLM to determine WHICH nodes fail given the cascade is triggered
prompt = f"""You are a financial network cascade analyzer. Analyze this network stress scenario.
NETWORK NODES: {nodes}
SHOCKED NODE: {shocked_node}
SHOCK MAGNITUDE: {shock_magnitude * 100:.0f}%
CONNECTIVITY STRENGTH: {connectivity_strength}
LIQUIDITY BUFFER LEVEL: {liquidity_buffer}
SYSTEMIC RISK SCORE: {systemic_risk * 100:.0f}%
CONGESTION LEVEL: {congestion_level}
CASCADE HAS BEEN TRIGGERED. Determine which nodes fail based on these rules:
1. The shocked node ALWAYS fails first
2. If shock > 70%, nodes most connected to the shocked node fail next
3. If connectivity > 1.5, failures spread to 50% of remaining nodes
4. If liquidity < 0.5, all nodes fail
5. Otherwise, 1-3 additional nodes fail proportional to shock magnitude
Respond with ONLY valid JSON (no markdown, no extra text):
{{"failed_nodes": ["NODE1.NS", "NODE2.NS"], "cascade_depth": 2, "analysis": "brief explanation"}}
IMPORTANT: Only include nodes from the NETWORK NODES list. The shocked node must be in failed_nodes."""
try:
response = llm_client.chat.completions.create(
model="llama-3.3-70b-versatile",
messages=[
{"role": "system", "content": "You are a financial risk analyst. Always respond with valid JSON only."},
{"role": "user", "content": prompt}
],
temperature=0.1, # Low temperature for deterministic output
max_tokens=500
)
result_text = response.choices[0].message.content.strip()
# Clean up response - remove markdown code blocks if present
if result_text.startswith("```"):
result_text = result_text.split("```")[1]
if result_text.startswith("json"):
result_text = result_text[4:]
result_text = result_text.strip()
result = json.loads(result_text)
failed_nodes = result.get("failed_nodes", [shocked_node])
# Ensure shocked node is in failed list
if shocked_node not in failed_nodes:
failed_nodes.insert(0, shocked_node)
# Filter to only valid nodes
failed_nodes = [n for n in failed_nodes if n in nodes]
# Generate failed edges (edges connected to failed nodes)
failed_edges = []
for fn in failed_nodes:
for other in nodes:
if other != fn and other not in failed_nodes:
failed_edges.append({"source": fn, "target": other})
return {
"status": "cascade",
"failed_nodes": failed_nodes,
"failed_edges": failed_edges,
"failure_ratio": len(failed_nodes) / len(nodes) if nodes else 0,
"cascade_depth": result.get("cascade_depth", 1),
"analysis": result.get("analysis", "Cascade triggered due to network stress.")
}
except Exception as e:
logger.error(f"LLM cascade analysis failed: {e}")
# Fallback: deterministic cascade based on rules
failed_nodes = [shocked_node]
if shock_magnitude > 0.8:
# Add ~50% of nodes
additional = int(len(nodes) * 0.5)
for n in nodes:
if n != shocked_node and len(failed_nodes) < additional + 1:
failed_nodes.append(n)
elif shock_magnitude > 0.6:
# Add 1-2 nodes
for n in nodes[:2]:
if n != shocked_node:
failed_nodes.append(n)
failed_edges = []
for fn in failed_nodes:
for other in nodes:
if other != fn and other not in failed_nodes:
failed_edges.append({"source": fn, "target": other})
return {
"status": "cascade",
"failed_nodes": failed_nodes,
"failed_edges": failed_edges,
"failure_ratio": len(failed_nodes) / len(nodes) if nodes else 0,
"cascade_depth": 1,
"analysis": f"Cascade triggered (fallback mode): {str(e)[:50]}"
}
# --- Configuration & Global State ---
app = FastAPI(title="Equilibrium Systemic Risk API")
# Add CORS middleware
app.add_middleware(
CORSMiddleware,
allow_origins=["http://localhost:3000", "http://127.0.0.1:3000"],
allow_credentials=True,
allow_methods=["*"],
allow_headers=["*"],
)
# Import and register routers
from routers import balance_sheet
app.include_router(balance_sheet.router, prefix="/api", tags=["balance_sheet"])
# Resolve paths relative to this file's directory
BASE_DIR = os.path.dirname(os.path.abspath(__file__))
CONFIG_PATH = os.path.join(BASE_DIR, "..", "frontend", "config.json")
MODEL_REPO = "Vansh180/Equilibrium-India-V1"
MODEL_FILENAME = "systemic_risk_model.keras"
SCALER_FILENAME = "scaler.pkl"
FEATURE_COLUMNS_FILENAME = "feature_columns.json"
# Global variables (loaded at startup)
config = {}
model = None
scaler = None
feature_columns = []
# --- Pydantic Models ---
class PredictionRequest(BaseModel):
tickers: List[str]
connectivity_scale: float = Field(default=1.0, gt=0, description="Scale factor for connectivity (must be > 0)")
liquidity_buffer_scale: float = Field(default=1.0, gt=0, description="Scale factor for liquidity buffer (must be > 0)")
class CCPFundsOutput(BaseModel):
initial_margin: float
variation_margin_flow: float
default_fund: float
ccp_capital: float
units: str = "indexed"
vm_is_flow: bool = True
class PredictionResponse(BaseModel):
predicted_next_systemic_risk: float
latest_S_t: float
latest_features: Dict[str, float]
used_tickers: List[str]
masked_tickers: List[str]
end_date: str
ccp_funds: CCPFundsOutput
# --- CCP Fund Computation ---
# Baseline constants (indexed units)
IM0 = 100
VM0 = 20
DF0 = 40
C0 = 10
# Reference values for normalization (typical observed ranges from training data)
# lambda_max typically ranges 1.5-3.5 for correlation matrices with CCP
# std_risk typically ranges 0.0-0.4 for normalized risk metrics
LAMBDA_REF = 2.5 # Reference lambda_max for normalization
STD_REF = 0.25 # Reference std_risk for normalization
def compute_ccp_funds(
systemic: float,
lambda_max: float,
std_risk: float,
connectivity_scale: float = 1.0,
liquidity_buffer_scale: float = 1.0
) -> Dict[str, Any]:
"""
Compute CCP fund requirements based on systemic risk and network metrics.
Args:
systemic: Predicted systemic risk score (0-1)
lambda_max: Maximum eigenvalue of adjacency matrix
std_risk: Standard deviation of risk across nodes
connectivity_scale: Scenario override for connectivity (default 1.0)
liquidity_buffer_scale: Scenario override for liquidity buffer (default 1.0)
Returns:
Dictionary with IM, VM, DF, CCP capital metrics
"""
# Normalize lambda_max and std_risk to 0-1 range using reference values
# Values above reference map to >1, below to <1
lambda_norm = min(1.0, lambda_max / LAMBDA_REF) if LAMBDA_REF > 0 else 0.0
std_norm = min(1.0, std_risk / STD_REF) if STD_REF > 0 else 0.0
# Compute CCP stress from normalized values, then clamp to [0,1]
# This ensures stress responds to changes rather than saturating
ccp_stress = 0.4 * lambda_norm + 0.3 * std_norm + 0.3 * systemic
ccp_stress = max(0.0, min(1.0, ccp_stress))
# Compute 4 values based on stress
im = IM0 * (1 + 1.5 * ccp_stress)
vm = VM0 * (1 + 2.0 * systemic) # VM is a flow based on systemic only
df = DF0 * (1 + 2.0 * max(0, ccp_stress - 0.5))
# CCP capital scales with stress: C_t = C_0 * (1 + k_C * CCPStress), clamped >= C_0
k_C = 9.0 # Scaling factor for range 10-100
ccp_capital = max(C0, C0 * (1 + k_C * ccp_stress))
# Apply scenario overrides
im *= connectivity_scale
df *= connectivity_scale
vm *= 1.0 / max(liquidity_buffer_scale, 1e-6) # Tight liquidity => bigger VM
return {
"initial_margin": round(im, 4),
"variation_margin_flow": round(vm, 4),
"default_fund": round(df, 4),
"ccp_capital": round(ccp_capital, 4),
"units": "indexed",
"vm_is_flow": True
}
def _run_ccp_funds_self_checks():
"""Run self-checks to validate compute_ccp_funds logic."""
# Check 1: If systemic increases, IM/VM should increase
low_sys = compute_ccp_funds(0.2, 0.5, 0.3)
high_sys = compute_ccp_funds(0.8, 0.5, 0.3)
assert high_sys["initial_margin"] >= low_sys["initial_margin"], "IM should increase with systemic"
assert high_sys["variation_margin_flow"] >= low_sys["variation_margin_flow"], "VM should increase with systemic"
# Check 2: DF should increase only when ccp_stress > 0.5
low_stress = compute_ccp_funds(0.1, 0.1, 0.1) # ccp_stress = 0.4*0.1 + 0.3*0.1 + 0.3*0.1 = 0.1
high_stress = compute_ccp_funds(0.8, 0.8, 0.8) # ccp_stress = 0.4*0.8 + 0.3*0.8 + 0.3*0.8 = 0.8
assert low_stress["default_fund"] == DF0, "DF should stay at baseline when ccp_stress <= 0.5"
assert high_stress["default_fund"] > DF0, "DF should increase when ccp_stress > 0.5"
# Check 3: Increasing connectivity_scale increases IM and DF
base = compute_ccp_funds(0.5, 0.5, 0.5, connectivity_scale=1.0)
scaled = compute_ccp_funds(0.5, 0.5, 0.5, connectivity_scale=1.5)
assert scaled["initial_margin"] > base["initial_margin"], "IM should increase with connectivity_scale"
assert scaled["default_fund"] > base["default_fund"], "DF should increase with connectivity_scale"
logger.info("CCP funds self-checks passed.")
# --- Helper Functions ---
def load_config():
"""Load configuration from config.json."""
global config
try:
# Try local config.json first (deployment)
if os.path.exists("config.json"):
config_file = "config.json"
else:
# Fallback for local development
config_file = "../frontend/config.json"
if os.path.exists(config_file):
with open(config_file, "r") as f:
config = json.load(f)
logger.info(f"Config loaded from {config_file}")
else:
# Fallback default config if file is missing (though unlikely in this setup)
logger.warning(f"Config file not found. Using hardcoded defaults.")
config = {
"tickers": ["HDFCBANK.NS", "KOTAKBANK.NS", "ICICIBANK.NS", "BAJFINANCE.NS", "BSE.NS",
"TCS.NS", "INFY.NS", "RELIANCE.NS", "SBIN.NS", "ADANIENT.NS",
"MRF.NS", "HINDUNILVR.NS", "TATASTEEL.NS", "AXISBANK.NS", "BHARTIARTL.NS"],
"ccp_name": "CCP",
"start": "2022-01-01",
"ret_window": 20,
"lookback": 20,
"delta_ccp": 0.1
}
except Exception as e:
logger.error(f"Error loading config: {e}")
raise
def setup_model_artifacts():
"""Download and load model, scaler, and feature columns."""
global model, scaler, feature_columns
try:
# 1. Download Model
logger.info(f"Downloading {MODEL_FILENAME} from {MODEL_REPO}...")
model_path = hf_hub_download(repo_id=MODEL_REPO, filename=MODEL_FILENAME, repo_type="model")
model = tf.keras.models.load_model(model_path) # type: ignore
logger.info("Model loaded successfully.")
# 2. Download Scaler
logger.info(f"Downloading {SCALER_FILENAME} from {MODEL_REPO}...")
scaler_path = hf_hub_download(repo_id=MODEL_REPO, filename=SCALER_FILENAME, repo_type="model")
scaler = joblib.load(scaler_path)
logger.info("Scaler loaded successfully.")
# 3. Download/Set Feature Columns
try:
logger.info(f"Attempting to download {FEATURE_COLUMNS_FILENAME}...")
cols_path = hf_hub_download(repo_id=MODEL_REPO, filename=FEATURE_COLUMNS_FILENAME, repo_type="model")
with open(cols_path, "r") as f:
feature_columns = json.load(f)
logger.info(f"Feature columns loaded: {feature_columns}")
except Exception as e:
logger.warning(f"Could not load feature_columns.json ({e}). Using default columns.")
feature_columns = ["lambda_max", "mean_risk", "max_risk", "std_risk", "S_lag1", "S_lag5"]
except Exception as e:
logger.critical(f"Failed to setup model artifacts: {e}")
raise RuntimeError("Model initialization failed.")
# --- Startup Event ---
@app.on_event("startup")
async def startup_event():
load_config()
setup_model_artifacts()
_run_ccp_funds_self_checks() # Validate CCP funds logic
# --- Core Logic ---
def compute_rolling_metrics(returns, window=20):
"""Compute rolling volatility and rolling max drawdown proxy."""
# Rolling Volatility
rolling_std = returns.rolling(window=window).std()
# Rolling Drawdown Proxy (simple)
# Using rolling max of price would be better, but we have returns.
# Let's reconstruct a cumulative return series for drawdown calculation roughly or use return based proxy.
# Standard practice with just returns:
# We need prices for standard drawdown. Let's assume we can use cumulative sum of log returns as log-price.
log_prices = returns.cumsum()
rolling_max = log_prices.rolling(window=window, min_periods=1).max()
drawdown = (log_prices - rolling_max) # This is log-drawdown, roughly % drawdown
# Invert so it's a positive risk metric? Drawdown is negative.
# Risk is usually magnitude. Let's take abs of drawdown (distance from peak).
rolling_drawdown = drawdown.abs()
return rolling_std, rolling_drawdown
def compute_features_for_subset(tickers_subset: List[str]):
"""
Main computational pipeline.
1. Fetch data for ALL config tickers.
2. Compute market-wide metrics (Adjacency, Eigenvector).
3. Apply masking: Set risk of unselected tickers to 0.
4. Compute Systemic Risk Payoff S_t.
5. Construct feature lags.
"""
all_tickers = config["tickers"]
start_date = config["start"]
ccp_name = config["ccp_name"]
delta_ccp = config["delta_ccp"]
lookback = config["lookback"]
# Identify indices
try:
# Create a boolean mask for selected tickers
# We need to maintain the order of 'all_tickers' for matrix operations
mask_vector = np.array([1.0 if t in tickers_subset else 0.0 for t in all_tickers])
except Exception as e:
raise HTTPException(status_code=400, detail=f"Error processing ticker subset: {e}")
# 1. Fetch Data
# We fetch data until today.
# yfinance auto_adjust=True
logger.info("Fetching market data...")
try:
raw_data = yf.download(all_tickers, start=start_date, progress=False, auto_adjust=True, threads=False)
if raw_data is None or raw_data.empty:
raise HTTPException(status_code=500, detail="No data returned from yfinance.")
data = raw_data['Close']
except HTTPException as he:
raise he
except Exception as e:
raise HTTPException(status_code=500, detail=f"Failed to fetch data from yfinance: {e}")
if data is None or data.empty:
raise HTTPException(status_code=500, detail="No Close price data returned from yfinance.")
# Handle single ticker case (though unlikely given list) causing Series instead of DataFrame
if isinstance(data, pd.Series):
data = data.to_frame()
# Reorder columns to match all_tickers list exactly (yfinance might sort them)
# Filter only columns that exist (in case some tickers failed)
existing_tickers = [t for t in all_tickers if t in data.columns]
data = data[existing_tickers]
# Recalculate mask based on existing tickers (if any were dropped by yfinance)
mask_vector = np.array([1.0 if t in tickers_subset and t in existing_tickers else 0.0 for t in existing_tickers])
N = len(existing_tickers)
# 2. Log Returns
returns = np.log(data / data.shift(1)).dropna()
# 3. Rolling Calculations
# We need to iterate over time to compute A_t, v_t, r_t, S_t
# Correlation window = 20
window = 20
# Lists to store time-series of metrics
s_t_series = []
lambda_max_series = []
# Store other risk metrics for feature creation
mean_risk_series = []
max_risk_series = []
std_risk_series = []
# Rolling Volatility & Drawdown (Vectorized)
rolling_std, rolling_dd = compute_rolling_metrics(returns, window)
# To compute min-max normalization, we need expanding window/historic min-max.
# Let's simplify and do expanding window from start of data.
# Combined Risk Metric: Volatility + Drawdown (Equal weight?)
# "rolling vol + rolling drawdown proxy" - Sum?
raw_risk = rolling_std + rolling_dd
# Normalize [0,1] expanding
# Note: expanding().min() / max() might be slow in loop, let's vectorise
expanding_min = raw_risk.expanding().min()
expanding_max = raw_risk.expanding().max()
# Avoid div by zero
denom = expanding_max - expanding_min
denom = denom.replace(0, 1.0) # Handle constant case
norm_risk = (raw_risk - expanding_min) / denom
# We can only compute calculating from `window` onwards
# And we need enough history for lags (lag 5 is max)
# We need to return exactly `lookback` (20) days of features.
# BUT, to compute S_lag5 for the *first* of those 20 days, we need S from 5 days before that.
# So we need to compute S_t for range: [end - lookback - 5, end]
required_history_len = lookback + 5
# Ensure we have enough data
if len(returns) < window + required_history_len:
raise HTTPException(status_code=400, detail=f"Insufficient data history. Need at least {window + required_history_len} days.")
# Slice the relevant period for iteration
# We usually want the "latest" prediction, so we process the tail.
# Let's process the last (required_history_len) days.
process_indices = returns.index[-required_history_len:]
# Pre-compute Rolling Correlations for efficiency?
# rolling(window).corr() returns a MultiIndex series.
# It might be heavy to do for all history. Let's do loop for just the needed days.
# History of S metrics to build lags
history_S = []
for date in process_indices:
# 1. Get correlation matrix for window ending at 'date'
# Data slice: date-window+1 to date
# returns.loc[:date].tail(window)
window_returns = returns.loc[:date].tail(window)
if len(window_returns) < window:
# Should not happen given logic above
s_t_series.append(0)
continue
# Correlation
corr_mat = window_returns.corr().values
# Fill NaNs (if constant price) with 0
corr_mat = np.nan_to_num(corr_mat)
# 2. Adjacency Matrix A
# Off-diagonal = max(0, corr)
A = np.maximum(0, corr_mat)
np.fill_diagonal(A, 0)
# 3. Add CCP
# A is N x N. New A is (N+1) x (N+1)
# Append column and row
# Column N: 0.1s
# Row N: 0.1s
# A[N,N] = 0
A_ext = np.zeros((N+1, N+1))
# Copy bank-bank block
A_ext[:N, :N] = A
# Add CCP edges
A_ext[:N, N] = delta_ccp # Bank -> CCP
A_ext[N, :N] = delta_ccp # CCP -> Bank
# 4. Compute Principal Eigenvector & Lambda Max
# Power iteration or linalg.eigh
# Since A is symmetric (corr is symmetric), eigh is good.
eigvals, eigvecs = np.linalg.eigh(A_ext)
# Max eigenvalue and vector
lambda_max = eigvals[-1]
v_t = eigvecs[:, -1]
# Ensure v_t is positive (Perron-Frobenius for non-negative matrices implies there's a non-negative eigenvector)
# Sometimes solver flips sign.
if np.sum(v_t) < 0:
v_t = -v_t
# 5. Node Risks r_t
# Get risk for this date
# norm_risk is a DataFrame, get row, convert to array
r_t_banks = norm_risk.loc[date].values
# Apply MASKING
# Set unselected tickers to 0
r_t_banks_masked = r_t_banks * mask_vector
# CCP Risk = 0
r_t_ext = np.append(r_t_banks_masked, 0.0)
# 6. Payoff S_t = r_t^T * A * v_t
# Dot product
# A * v
Av = np.dot(A_ext, v_t)
# r * Av
S_t = np.dot(r_t_ext, Av)
# 7. Statistics for Features
# Compute stats on the MASKED banks risk (excluding CCP)
# "mean_risk, max_risk, std_risk"
# Using masked values might skew mean to 0 if many are masked.
# But this reflects the "effective" system state if unselected are removed.
# Let's use the masked vectors.
# Note: If we really masked checks (set to 0), maybe we should exclude them from mean/std?
# But for fixed feature vector size, usually we just compute on the vector.
mean_r = np.mean(r_t_banks_masked)
max_r = np.max(r_t_banks_masked)
std_r = np.std(r_t_banks_masked)
# Store
history_S.append(S_t)
lambda_max_series.append(lambda_max)
mean_risk_series.append(mean_r)
max_risk_series.append(max_r)
std_risk_series.append(std_r)
# Convert history to DataFrame to build features
feature_df = pd.DataFrame({
"lambda_max": lambda_max_series,
"mean_risk": mean_risk_series,
"max_risk": max_risk_series,
"std_risk": std_risk_series,
"S_t": history_S
}, index=process_indices)
# Create Lags
feature_df["S_lag1"] = feature_df["S_t"].shift(1)
feature_df["S_lag5"] = feature_df["S_t"].shift(5)
# Drop NaNs created by shifting
# We calculated `lookback + 5` days.
# Shifting by 5 loses first 5.
feature_df = feature_df.dropna()
# Select last `lookback` (20) rows
feature_df = feature_df.tail(lookback)
if len(feature_df) < lookback:
logger.error(f"Not enough data after lag creation. Have {len(feature_df)}, need {lookback}")
raise HTTPException(status_code=400, detail="Insufficient data for feature window.")
# Select feature columns in correct order
# Ensure columns match model expectation
final_features = feature_df[feature_columns]
return final_features, existing_tickers
# --- Endpoint ---
@app.post("/predict", response_model=PredictionResponse)
async def predict(request: PredictionRequest):
if not model or not scaler:
raise HTTPException(status_code=503, detail="Model not loaded.")
# Input validation
tickers_subset = request.tickers
if not tickers_subset:
raise HTTPException(status_code=400, detail="Tickers list cannot be empty.")
valid_tickers = set(config["tickers"])
invalid = [t for t in tickers_subset if t not in valid_tickers]
if invalid:
raise HTTPException(status_code=400, detail=f"Invalid tickers: {invalid}. Must be in config.")
# Compute Features
try:
features_df, available_tickers = compute_features_for_subset(tickers_subset)
except HTTPException as he:
raise he
except Exception as e:
logger.error(f"Computation error: {e}", exc_info=True)
raise HTTPException(status_code=500, detail=str(e))
# Scale Features
X = features_df.values
try:
X_scaled = scaler.transform(X)
except Exception as e:
logger.error(f"Scaling error: {e}")
raise HTTPException(status_code=500, detail=f"Scaling failed: {e}")
# Reshape for LSTM/GRU: (1, 20, 6)
# features_df should have 20 rows
X_reshaped = X_scaled.reshape(1, config["lookback"], len(feature_columns))
# Predict
try:
prediction = model.predict(X_reshaped)
risk_score = float(prediction[0][0])
# Cap at 1.0 (100%) to prevent exceeding maximum
risk_score = min(risk_score, 1.0)
except Exception as e:
logger.error(f"Prediction error: {e}")
raise HTTPException(status_code=500, detail=f"Model prediction failed: {e}")
# Latest Data
latest_row = features_df.iloc[-1]
last_date = features_df.index[-1].strftime("%Y-%m-%d")
# Extract metrics for CCP funds computation
# lambda_max and std_risk from latest features
lambda_max_val = float(latest_row.get("lambda_max", 0.0))
std_risk_val = float(latest_row.get("std_risk", 0.0))
# Compute CCP funds
ccp_funds = compute_ccp_funds(
systemic=risk_score,
lambda_max=lambda_max_val,
std_risk=std_risk_val,
connectivity_scale=request.connectivity_scale,
liquidity_buffer_scale=request.liquidity_buffer_scale
)
return {
"predicted_next_systemic_risk": risk_score,
"latest_S_t": latest_row.get("S_t", 0.0) if "S_t" in latest_row else 0.0,
"latest_features": latest_row.to_dict(),
"used_tickers": available_tickers,
"masked_tickers": [t for t in config["tickers"] if t not in tickers_subset],
"end_date": last_date,
"ccp_funds": ccp_funds
}
@app.get("/health")
def health():
return {"status": "ok", "model_loaded": model is not None}
# --- Simulation Models ---
class SimulationRequest(BaseModel):
tickers: List[str]
start: Optional[str] = None # Optional start date, default from config
end: Optional[str] = None # Optional end date, default latest
shocked_node: str # Must be in tickers universe
shock_magnitude: float = Field(default=0.3, ge=0, le=1) # 0..1
connectivity_strength: float = Field(default=1.0, ge=0.1, le=2.0)
liquidity_buffer_level: float = Field(default=1.0, ge=0.1, le=2.0)
ccp_strictness: float = Field(default=0.5, ge=0, le=1)
correlation_regime: float = Field(default=0.7, ge=0, le=1)
steps: int = Field(default=10, ge=1, le=50) # Cascade steps
class CongestionOutput(BaseModel):
level: str # "low", "medium", "high"
most_congested_node: str
max_score: float
class CascadeOutput(BaseModel):
status: str # "stable" or "cascade"
failed_nodes: List[str]
failed_edges: List[Dict[str, str]] # [{"source": "A.NS", "target": "B.NS"}]
failure_ratio: float
cascade_depth: int
analysis: str # LLM explanation
class SimulationResponse(BaseModel):
end_date: str
used_tickers: List[str]
masked_tickers: List[str]
latest_features: Dict[str, float]
latest_payoff_S: float
predicted_next_systemic_risk: float
congestion: CongestionOutput
cascade: CascadeOutput
ccp_funds: CCPFundsOutput
# --- Simulation Endpoint ---
@app.post("/simulate", response_model=SimulationResponse)
async def simulate(request: SimulationRequest):
"""
Run a network stress simulation with node shocks and global controls.
"""
if not model or not scaler:
raise HTTPException(status_code=503, detail="Model not loaded.")
# NEW: Allow dynamic tickers beyond just config.json
all_tickers = list(set(config["tickers"] + request.tickers))
valid_tickers = set(all_tickers)
tickers_subset = [t for t in request.tickers if t] # Take all requested
if not tickers_subset:
tickers_subset = all_tickers
logger.info("No tickers provided. Falling back to default universe.")
# Validate shocked_node
shocked_node = request.shocked_node
if shocked_node not in valid_tickers:
logger.warning(f"shocked_node '{shocked_node}' not in allowed universe. Using '{tickers_subset[0]}' as quantitative proxy.")
shocked_node = tickers_subset[0]
# Fetch data
start_date = request.start if request.start else config["start"]
delta_ccp = config["delta_ccp"]
ret_window = config.get("ret_window", 20)
lookback = config.get("lookback", 20)
try:
end_param = request.end if request.end else None
raw_data = yf.download(all_tickers, start=start_date, end=end_param, progress=False, auto_adjust=True, threads=False)
if raw_data is None or raw_data.empty:
raise HTTPException(status_code=500, detail="No data returned from yfinance.")
data = raw_data['Close']
except HTTPException as he:
raise he
except Exception as e:
raise HTTPException(status_code=500, detail=f"Failed to fetch data: {e}")
if isinstance(data, pd.Series):
data = data.to_frame()
existing_tickers = [t for t in all_tickers if t in data.columns]
data = data[existing_tickers]
N = len(existing_tickers)
# Log returns
returns = np.log(data / data.shift(1)).dropna()
if len(returns) < ret_window:
raise HTTPException(status_code=400, detail=f"Insufficient data. Need at least {ret_window} days.")
# Use latest window for correlation
latest_returns = returns.tail(ret_window)
end_date = returns.index[-1].strftime("%Y-%m-%d")
# --- Step 1: Build baseline adjacency A ---
corr_mat = latest_returns.corr().values
corr_mat = np.nan_to_num(corr_mat)
A = np.maximum(0, corr_mat)
np.fill_diagonal(A, 0)
# --- Step 2: Add CCP edges ---
# CCP strictness modifies edge weight
ccp_edge = delta_ccp * (1 + request.ccp_strictness)
# Extend matrix for CCP
A_ext = np.zeros((N+1, N+1))
A_ext[:N, :N] = A
A_ext[:N, N] = ccp_edge # Bank -> CCP
A_ext[N, :N] = ccp_edge # CCP -> Bank
# --- Step 3: Apply global sliders ---
# Connectivity strength
A_ext = request.connectivity_strength * A_ext
# Correlation regime (density gating)
tau = 1 - request.correlation_regime
A_ext[A_ext < tau] = 0
# --- Step 4: Build node risk vector ---
# Compute rolling risk metrics
rolling_std = returns.rolling(window=ret_window).std()
log_prices = returns.cumsum()
rolling_max = log_prices.rolling(window=ret_window, min_periods=1).max()
drawdown = (log_prices - rolling_max).abs()
raw_risk = rolling_std + drawdown
# Get latest risk
latest_risk = raw_risk.iloc[-1].values
# Normalize to [0,1]
risk_min = np.nanmin(latest_risk)
risk_max = np.nanmax(latest_risk)
if risk_max > risk_min:
r = (latest_risk - risk_min) / (risk_max - risk_min)
else:
r = np.zeros(N)
r = np.nan_to_num(r, nan=0.0)
# Create mask for selected tickers
mask_vector = np.array([1.0 if t in tickers_subset else 0.0 for t in existing_tickers])
# Apply masking
r_masked = r * mask_vector
# Apply shock
if request.shocked_node in existing_tickers:
shock_idx = existing_tickers.index(request.shocked_node)
r_masked[shock_idx] = np.clip(r_masked[shock_idx] + request.shock_magnitude, 0, 1)
# Extend with CCP (risk = 0)
r_ext = np.append(r_masked, 0.0)
mask_ext = np.append(mask_vector, 0.0) # CCP excluded from payoff
# --- Step 5: Eigen computations ---
eigvals, eigvecs = np.linalg.eigh(A_ext)
lambda_max = float(eigvals[-1])
v = eigvecs[:, -1]
if np.sum(v) < 0:
v = -v
v = v / (np.linalg.norm(v) + 1e-9) # Normalize
# --- Step 6: Payoff and congestion ---
# S = r^T A (m ⊙ v)
masked_v = mask_ext * v
Av = np.dot(A_ext, masked_v)
S = float(np.dot(r_ext, Av))
# Congestion scores - ONLY for selected tickers
# Buffer baseline: 1 - r, scaled by liquidity
b = (1 - r_ext) * request.liquidity_buffer_level
b = np.maximum(b, 0.1) # Minimum buffer
# Congestion score per node (only for selected nodes)
congestion_scores = np.zeros(N+1)
selected_indices = [i for i in range(N) if mask_vector[i] == 1]
for i in range(N+1):
if i < N and mask_vector[i] == 1: # Only compute for selected nodes
influence = np.dot(A_ext[:, i], r_ext)
congestion_scores[i] = (v[i] * influence) / (b[i] + 1e-6)
else:
congestion_scores[i] = -np.inf # Ignore non-selected nodes
# Find most congested among SELECTED nodes only
if len(selected_indices) > 0:
selected_congestion = [congestion_scores[i] for i in selected_indices]
max_congestion = float(np.max(selected_congestion))
most_congested_idx = selected_indices[int(np.argmax(selected_congestion))]
most_congested_node = existing_tickers[most_congested_idx]
else:
max_congestion = 0.0
most_congested_node = "None"
if max_congestion < 1:
congestion_level = "low"
elif max_congestion < 1.5:
congestion_level = "medium"
else:
congestion_level = "high"
# --- Step 7: ML feature row and prediction ---
mean_risk = float(np.mean(r_masked))
max_risk = float(np.max(r_masked))
std_risk = float(np.std(r_masked))
# Features in order
feature_row = {
"lambda_max": lambda_max,
"mean_risk": mean_risk,
"max_risk": max_risk,
"std_risk": std_risk,
"S_lag1": S, # Approximate for simulation
"S_lag5": S # Approximate for simulation
}
# Create lookback sequence (repeat feature row)
feature_array = np.array([[feature_row.get(col, 0) for col in feature_columns] for _ in range(lookback)])
# Scale
try:
X_scaled = scaler.transform(feature_array)
except Exception as e:
raise HTTPException(status_code=500, detail=f"Scaling failed: {e}")
X_reshaped = X_scaled.reshape(1, lookback, len(feature_columns))
# Predict
try:
prediction = model.predict(X_reshaped, verbose=0)
risk_score = float(prediction[0][0])
# Cap at 1.0 (100%) to prevent exceeding maximum
risk_score = min(risk_score, 1.0)
except Exception as e:
raise HTTPException(status_code=500, detail=f"Model prediction failed: {e}")
# --- Step 8: CCP Funds calculation ---
ccp_funds = compute_ccp_funds(
systemic=risk_score,
lambda_max=lambda_max,
std_risk=std_risk,
connectivity_scale=request.connectivity_strength,
liquidity_buffer_scale=request.liquidity_buffer_level
)
# --- Step 9: LLM-Based Cascade Analysis ---
# Use LLM with hard boundaries to determine cascade
used_tickers_list = [t for t in existing_tickers if t in tickers_subset]
cascade_result = analyze_cascade_with_llm(
nodes=used_tickers_list,
shocked_node=request.shocked_node,
shock_magnitude=request.shock_magnitude,
connectivity_strength=request.connectivity_strength,
liquidity_buffer=request.liquidity_buffer_level,
systemic_risk=risk_score,
congestion_level=congestion_level
)
return {
"end_date": end_date,
"used_tickers": used_tickers_list,
"masked_tickers": [t for t in all_tickers if t not in tickers_subset],
"latest_features": feature_row,
"latest_payoff_S": S,
"predicted_next_systemic_risk": risk_score,
"congestion": {
"level": congestion_level,
"most_congested_node": most_congested_node,
"max_score": round(max_congestion, 4)
},
"cascade": cascade_result,
"ccp_funds": ccp_funds
}
# --- News Generation Endpoints ---
class NewsGenerateRequest(BaseModel):
tickers: List[str]
count: int = Field(default=10, ge=1, le=20)
class NewsItem(BaseModel):
id: str
ticker: str
company_name: str
headline: str
summary: str
sentiment: str # positive, negative, neutral
confidence: float # 0-1
timestamp: str
is_breaking: bool = False
class NewsGenerateResponse(BaseModel):
news: List[NewsItem]
class NewsDefaultRequest(BaseModel):
ticker: str
company_name: str
default_magnitude: float = Field(ge=0, le=1)
class NewsDefaultResponse(BaseModel):
news: NewsItem
def generate_news_with_llm(tickers: List[str], count: int, is_default: bool = False, default_ticker: str = None, default_company: str = None, default_magnitude: float = 0.0) -> List[Dict[str, Any]]:
"""Generate realistic financial news using LLM with sentiment analysis."""
# Get company names from tickers
ticker_to_name = {}
for ticker in tickers:
try:
stock = yf.Ticker(ticker)
info = stock.info
ticker_to_name[ticker] = info.get('shortName', info.get('longName', ticker.replace('.NS', '')))
except:
ticker_to_name[ticker] = ticker.replace('.NS', '')
if is_default:
# Generate breaking default news
prompt = f"""Generate a realistic breaking financial news headline and 2-sentence summary about {default_company} ({default_ticker}) facing a severe default/liquidity crisis in the Indian market.
The news should be dramatic and reflect a severity level of {int(default_magnitude * 100)}%.
Respond in this exact JSON format:
{{
"headline": "BREAKING: [Urgent headline about default/crisis]",
"summary": "[2-sentence detailed summary explaining the situation]",
"sentiment": "negative",
"confidence": [0.85-0.99]
}}"""
else:
# Generate regular market news
companies_str = ", ".join([f"{name} ({ticker})" for ticker, name in list(ticker_to_name.items())[:5]])
prompt = f"""Generate {count} realistic financial news headlines for Indian companies: {companies_str}.
Create diverse news covering:
- Quarterly earnings (positive/negative)
- Market movements
- Business expansions
- Regulatory updates
- Industry trends
For EACH news item, classify sentiment as positive, negative, or neutral with a confidence score (0-1).
Respond in this exact JSON array format:
[
{{
"ticker": "TCS.NS",
"headline": "[Realistic headline]",
"summary": "[1-2 sentence summary]",
"sentiment": "positive|negative|neutral",
"confidence": 0.XX
}}
]"""
try:
response = llm_client.chat.completions.create(
model="llama-3.3-70b-versatile",
messages=[
{"role": "system", "content": "You are a financial news generator for Indian markets. Generate realistic, professional news in JSON format only."},
{"role": "user", "content": prompt}
],
temperature=0.8,
max_tokens=1500
)
content = response.choices[0].message.content.strip()
# Extract JSON from response
if '```json' in content:
content = content.split('```json')[1].split('```')[0].strip()
elif '```' in content:
content = content.split('```')[1].split('```')[0].strip()
news_data = json.loads(content)
if is_default:
# Single breaking news item
return [{
"ticker": default_ticker,
"company_name": default_company,
"headline": news_data["headline"],
"summary": news_data["summary"],
"sentiment": "negative",
"confidence": news_data.get("confidence", 0.95),
"is_breaking": True
}]
else:
# Multiple news items
news_items = []
for item in news_data[:count]:
ticker = item.get("ticker", tickers[0])
news_items.append({
"ticker": ticker,
"company_name": ticker_to_name.get(ticker, ticker.replace('.NS', '')),
"headline": item["headline"],
"summary": item.get("summary", ""),
"sentiment": item["sentiment"],
"confidence": item.get("confidence", 0.75),
"is_breaking": False
})
return news_items
except Exception as e:
logger.error(f"News generation failed: {e}")
# Fallback to template news
if is_default:
return [{
"ticker": default_ticker,
"company_name": default_company,
"headline": f"BREAKING: {default_company} Faces Severe Liquidity Crisis",
"summary": f"{default_company} has declared inability to meet short-term obligations amid market stress. Regulators have been notified.",
"sentiment": "negative",
"confidence": 0.90,
"is_breaking": True
}]
else:
return [
{
"ticker": ticker,
"company_name": ticker_to_name.get(ticker, ticker.replace('.NS', '')),
"headline": f"{ticker_to_name.get(ticker, ticker.replace('.NS', ''))} Reports Stable Quarter",
"summary": "Company performance in line with market expectations.",
"sentiment": "neutral",
"confidence": 0.70,
"is_breaking": False
}
for ticker in tickers[:count]
]
@app.post("/news/generate", response_model=NewsGenerateResponse)
def generate_news(request: NewsGenerateRequest):
"""Generate realistic financial news feed for given tickers."""
news_data = generate_news_with_llm(request.tickers, request.count)
news_items = []
for idx, item in enumerate(news_data):
news_items.append(NewsItem(
id=f"news-{datetime.now().timestamp()}-{idx}",
ticker=item["ticker"],
company_name=item["company_name"],
headline=item["headline"],
summary=item["summary"],
sentiment=item["sentiment"],
confidence=item["confidence"],
timestamp=datetime.now().isoformat(),
is_breaking=False
))
return NewsGenerateResponse(news=news_items)
@app.post("/news/default", response_model=NewsDefaultResponse)
def generate_default_news(request: NewsDefaultRequest):
"""Generate breaking news for bank default event."""
news_data = generate_news_with_llm(
[request.ticker],
1,
is_default=True,
default_ticker=request.ticker,
default_company=request.company_name,
default_magnitude=request.default_magnitude
)
item = news_data[0]
return NewsDefaultResponse(
news=NewsItem(
id=f"breaking-{datetime.now().timestamp()}",
ticker=item["ticker"],
company_name=item["company_name"],
headline=item["headline"],
summary=item["summary"],
sentiment="negative",
confidence=item["confidence"],
timestamp=datetime.now().isoformat(),
is_breaking=True
)
)
|