MacroLens / code /dataloader /canonical_indices.py
itouchz's picture
Duplicate from macrolens/MacroLens
ff4becd
Raw
History Blame Contribute Delete
23.1 kB
"""Canonical-indices generator.
Reads the immutable benchmark artifacts and produces deterministic,
stratified train/eval index lists per task. Every family runner reads
from here so cross-method comparison is fair: same N, same instances,
same indices.
The cache lives under `experiments/cache/canonical_indices/<key>/`
where `<key>` encodes (n_eval, n_train, seed, stratifier_version) -- so
changing any sample-budget parameter produces a new cache directory.
The cache is an experiment-side speed optimisation; the canonical
dataset tree under `data_small_caps/` contains only raw and derived
benchmark artifacts (which are immutable).
Per-task stratification:
T1 (TSF): sector x market_cap_quartile -> (ticker, anchor_date)
T2 (Val-PT): full 30% ticker holdout -> (ticker, date)
T3 (Stmt-Gen): per-(ticker, fiscal_year) holdout -> (ticker, fiscal_year)
T4 (Scen-Ret): sector x mcap_q x event_type -> (scenario_id, ticker)
T5 (Val-Priv): full 30% ticker holdout -> (ticker, date)
T6 (Gen-Eval): per-(ticker, fiscal_year) holdout -> (ticker, fiscal_year)
T7 (RE-Val): property_type x state -> address
"""
from __future__ import annotations
import logging
from pathlib import Path
from typing import Literal
import numpy as np
import pandas as pd
from .. import config
from . import budgets
from .budgets import EVAL_N_PER_TASK, TRAIN_N_PER_TASK, SEED, Task
logger = logging.getLogger(__name__)
Split = Literal["train", "eval"]
def _cfg_get_lookback(granularity: str) -> int:
"""Return the canonical (shortest) lookback for ``granularity``."""
return config.get_lookback_windows(granularity)[0] if hasattr(config, "get_lookback_windows") else 63
def _provenance_suffix(granularity: str) -> str:
"""16-char SHA-256-derived suffix encoding the relevant benchmark
parquets for ``granularity``. Mutating any of those parquets changes
the cache key, forcing canonical-indices regeneration.
"""
from ._provenance import sha256_combined
bench_dir = config.get_benchmark_dir(granularity)
candidates = [
bench_dir / "panel_train.parquet",
bench_dir / "panel_test.parquet",
bench_dir / "valuation_inputs.parquet",
bench_dir / "valuation_ground_truth.parquet",
bench_dir / "private_valuation_inputs.parquet",
bench_dir / "private_valuation_ground_truth.parquet",
bench_dir / "generation_inputs.parquet",
bench_dir / "generation_ground_truth.parquet",
bench_dir / "generator_eval_inputs.parquet",
bench_dir / "generator_eval_ground_truth.parquet",
bench_dir / "scenario_forecast_ground_truth.parquet",
bench_dir / "scenarios.parquet",
bench_dir / "re_train_properties.parquet",
bench_dir / "re_eval_inputs.parquet",
bench_dir / "re_eval_ground_truth.parquet",
]
existing = [p for p in candidates if p.exists()]
if not existing:
return "noprov"
return sha256_combined(existing)
def _cache_dir(granularity: str, key: str | None = None) -> Path:
"""Return the cache directory for a given budget key.
Lives under ``experiments/cache/canonical_indices/`` (experiment-side
speed optimisation, regenerable on miss). The canonical dataset tree
under ``data_small_caps/`` contains only raw and derived benchmark
artifacts; experiment-side caches NEVER live there.
The cache key suffix encodes (i) a SHA-256 over the benchmark
parquets and (ii) the current ``max(lookback)`` and ``max(horizon)``
for ``granularity``. Either upstream-data drift or a horizon/lookback
config change atomically invalidates the cache.
"""
k = key or budgets.cache_key()
suffix = _provenance_suffix(granularity)
max_lb = max(config.get_lookback_windows(granularity))
max_h = max(config.get_horizons(granularity))
# Resolve experiments/ as a sibling of dataloader/ (this file lives
# at projects/.../whatif_bench/dataloader/canonical_indices.py).
experiments_dir = Path(__file__).resolve().parents[1] / "experiments"
return (
experiments_dir
/ "cache"
/ "canonical_indices"
/ granularity
/ f"{k}_prov={suffix}_lb={max_lb}_h={max_h}"
)
def _stratified_sample(
df: pd.DataFrame,
n: int,
strata_cols: list[str],
seed: int,
) -> pd.DataFrame:
"""Stratified subsample of `df` to size `n`, preserving the joint
distribution of `strata_cols` (Cartesian-product strata, with
proportional allocation and remainder spread by row order).
Deterministic at fixed `seed`. If `n >= len(df)`, returns df shuffled.
"""
if n >= len(df):
return df.sample(frac=1.0, random_state=seed).reset_index(drop=True)
# Drop rows with NaN in any stratifier column -- they would form a
# spurious "missing" stratum.
valid_mask = df[strata_cols].notna().all(axis=1)
df_valid = df[valid_mask].copy()
if df_valid.empty:
# Fall back to uniform random
return df.sample(n=n, random_state=seed).reset_index(drop=True)
df_valid["_stratum"] = df_valid[strata_cols].astype(str).agg("|".join, axis=1)
rng = np.random.RandomState(seed)
out_rows: list[pd.DataFrame] = []
total = len(df_valid)
for stratum, grp in df_valid.groupby("_stratum"):
# Proportional allocation; at least 1 if stratum has rows.
q = max(1, round(len(grp) * n / total))
q = min(q, len(grp))
out_rows.append(grp.sample(n=q, random_state=rng.randint(0, 2**31 - 1)))
out = pd.concat(out_rows, ignore_index=True)
# Trim or top-up to exactly n
if len(out) > n:
out = out.sample(n=n, random_state=seed).reset_index(drop=True)
elif len(out) < n:
# Top up with non-selected rows (still stratified by selection above)
remaining = df_valid.loc[~df_valid.index.isin(out.index)]
extra = remaining.sample(n=min(n - len(out), len(remaining)), random_state=seed)
out = pd.concat([out, extra], ignore_index=True)
return out.drop(columns=["_stratum"]).reset_index(drop=True)
# ── Per-task generators ───────────────────────────────────────────────────
def _gen_t1(
granularity: str,
n_eval: int,
n_train: int,
seed: int,
) -> dict[Split, pd.DataFrame]:
"""T1 TSF: stratified by sector x market_cap_quartile.
Each (ticker, anchor_date) pair must admit a complete
``(lookback, max_horizon)`` window inside the corresponding split's panel,
so every horizon evaluated by every T1 runner reuses the same anchor set.
Concretely, for a ticker with ``T`` panel rows we keep only anchor dates
at per-ticker positions ``[lookback, T - max_horizon - 1]``.
Returns DataFrames with columns (ticker, anchor_date, sector, mcap_q).
"""
from .. import config as _cfg
# Build the canonical anchor pool against the SHORTEST lookback and
# the LONGEST horizon. The test panel (post-2024-09-03) is ~378
# trading days; pairing max_lookback (252) with max_horizon (252)
# exhausts it. Methods that want a longer lookback can request it
# at load time (load(..., lookback=252)); anchors with insufficient
# history will be dropped by _build_t1_x_y and counted in
# ``meta.attrs["n_canonical_dropped"]``.
lookback = _cfg.get_lookback_windows(granularity)[0]
max_horizon = max(_cfg.get_horizons(granularity))
bench_dir = config.get_benchmark_dir(granularity)
train = pd.read_parquet(
bench_dir / "panel_train.parquet",
columns=["ticker", "date", "sector", "derived_market_cap"],
)
test = pd.read_parquet(
bench_dir / "panel_test.parquet",
columns=["ticker", "date", "sector", "derived_market_cap"],
)
# Latest market_cap per ticker for the quartile assignment (across the
# full panel, so train and eval split on the same definition).
latest = (
pd.concat([train, test], ignore_index=True)
.sort_values("date")
.groupby("ticker")
.tail(1)[["ticker", "derived_market_cap"]]
)
latest["mcap_q"] = pd.qcut(
latest["derived_market_cap"].clip(lower=1),
4, labels=["Q1", "Q2", "Q3", "Q4"], duplicates="drop",
)
mcap_q = dict(zip(latest["ticker"], latest["mcap_q"]))
def _restrict_to_valid_anchors(df: pd.DataFrame) -> pd.DataFrame:
"""Keep only rows at per-ticker positions [lookback, T-max_horizon-1]
so a complete (lookback + max_horizon) window fits."""
df = df.sort_values(["ticker", "date"]).reset_index(drop=True)
df["_row_in_ticker"] = df.groupby("ticker", sort=False).cumcount()
df["_ticker_len"] = df.groupby("ticker", sort=False)["date"].transform("size")
valid = (df["_row_in_ticker"] >= lookback) & (
df["_row_in_ticker"] < df["_ticker_len"] - max_horizon
)
return df.loc[valid].drop(columns=["_row_in_ticker", "_ticker_len"])
out: dict[Split, pd.DataFrame] = {}
for split, df in (("train", train), ("eval", test)):
df = _restrict_to_valid_anchors(df)
df = df.rename(columns={"date": "anchor_date"}).copy()
df["mcap_q"] = df["ticker"].map(mcap_q)
n = n_train if split == "train" else n_eval
sampled = _stratified_sample(df, n, ["sector", "mcap_q"], seed)
out[split] = sampled[["ticker", "anchor_date", "sector", "mcap_q"]]
return out
def _gen_t2_t5(
granularity: str,
task: str,
n_eval: int,
n_train: int,
seed: int,
) -> dict[Split, pd.DataFrame]:
"""T2 (Val-PT) and T5 (Val-Priv): full 30% ticker holdout for eval;
latest snapshot per non-holdout ticker for train.
Returns DataFrames with columns (ticker, date).
"""
bench_dir = config.get_benchmark_dir(granularity)
if task == "T2":
inputs_path = bench_dir / "valuation_inputs.parquet"
gt_path = bench_dir / "valuation_ground_truth.parquet"
else:
inputs_path = bench_dir / "private_valuation_inputs.parquet"
gt_path = bench_dir / "private_valuation_ground_truth.parquet"
inputs = pd.read_parquet(inputs_path, columns=["ticker", "date", "sector"])
gt = pd.read_parquet(gt_path, columns=["ticker", "date"])
# Restrict the eval pool to (ticker, date) pairs that are present in
# BOTH inputs and gt. Without this, ~21 quarter-end snapshots per
# task have inputs but no gt (close or shares_outstanding missing
# on that date), and the loader silently dropped them at merge
# time so canonical-eval N came up short of the budget.
inputs["date"] = pd.to_datetime(inputs["date"])
gt["date"] = pd.to_datetime(gt["date"])
eval_pool = inputs.merge(gt, on=["ticker", "date"], how="inner")
train_panel = pd.read_parquet(
bench_dir / "panel_train.parquet",
columns=["ticker", "date", "sector"],
)
holdout_tickers = set(inputs["ticker"].unique())
non_holdout = train_panel[~train_panel["ticker"].isin(holdout_tickers)]
# Latest snapshot per non-holdout ticker as the train set.
train_latest = (
non_holdout.sort_values("date").groupby("ticker").tail(1)
.reset_index(drop=True)
)
rng = np.random.RandomState(seed)
train_idx = train_latest.sample(
n=min(n_train, len(train_latest)), random_state=rng.randint(0, 2**31 - 1),
).reset_index(drop=True)
eval_idx = eval_pool.sample(
n=min(n_eval, len(eval_pool)), random_state=rng.randint(0, 2**31 - 1),
).reset_index(drop=True)
return {"train": train_idx, "eval": eval_idx}
def _gen_t3_t6(
granularity: str,
task: str,
n_eval: int,
n_train: int,
seed: int,
) -> dict[Split, pd.DataFrame]:
"""T3 (Stmt-Gen) and T6 (Gen-Eval): per-(ticker, fiscal_year) split.
Every ticker in the per-task ground-truth file is also in the
holdout (`*_inputs.parquet` lists holdout tickers only), so a
plain ``~ticker.isin(holdout)`` train pool would always be empty.
Instead: per ticker, the **latest** fiscal year is the eval anchor
and **earlier** fiscal years are train anchors. Train-eval are
cleanly separated by fiscal year within ticker; both pools are
non-empty as long as a ticker has >=2 reported fiscal years.
Eval = unique (ticker, latest_fiscal_year) pairs across all tickers.
Train = unique (ticker, prior_fiscal_year) pairs across all tickers.
"""
bench_dir = config.get_benchmark_dir(granularity)
if task == "T3":
gt_path = bench_dir / "generation_ground_truth.parquet"
else:
gt_path = bench_dir / "generator_eval_ground_truth.parquet"
gt = pd.read_parquet(gt_path)
if "fiscal_year" not in gt.columns:
if "filing_date" in gt.columns:
gt["fiscal_year"] = pd.to_datetime(gt["filing_date"]).dt.year
else:
gt["fiscal_year"] = 0
pairs = gt[["ticker", "fiscal_year"]].drop_duplicates().reset_index(drop=True)
pairs["fiscal_year"] = pd.to_numeric(pairs["fiscal_year"], errors="coerce")
pairs = pairs.dropna(subset=["fiscal_year"]).copy()
pairs["fiscal_year"] = pairs["fiscal_year"].astype(int)
# Per-ticker: latest FY -> eval, earlier FYs -> train
pairs = pairs.sort_values(["ticker", "fiscal_year"]).reset_index(drop=True)
pairs["_rank_desc"] = pairs.groupby("ticker")["fiscal_year"].rank(
method="first", ascending=False,
)
eval_pairs = pairs[pairs["_rank_desc"] == 1][["ticker", "fiscal_year"]]
train_pairs = pairs[pairs["_rank_desc"] > 1][["ticker", "fiscal_year"]]
rng = np.random.RandomState(seed)
eval_idx = eval_pairs.sample(
n=min(n_eval, len(eval_pairs)), random_state=rng.randint(0, 2**31 - 1),
).reset_index(drop=True)
train_idx = train_pairs.sample(
n=min(n_train, len(train_pairs)), random_state=rng.randint(0, 2**31 - 1),
).reset_index(drop=True)
return {"train": train_idx, "eval": eval_idx}
def _gen_t4(
granularity: str,
n_eval: int,
n_train: int,
seed: int,
) -> dict[Split, pd.DataFrame]:
"""T4 Scen-Ret: stratified by sector x mcap_q x event_type.
Returns DataFrames with columns (scenario_id, ticker, event_type).
"""
bench_dir = config.get_benchmark_dir(granularity)
gt = pd.read_parquet(
bench_dir / "scenario_forecast_ground_truth.parquet",
columns=["scenario_id", "ticker", "event_type", "event_date",
"actual_return_pct"],
)
gt = gt.dropna(subset=["actual_return_pct"])
# Use the panel-train cutoff as the train/eval split anchor (matches
# the canonical T1 split semantics).
panel_train_df = pd.read_parquet(
bench_dir / "panel_train.parquet", columns=["ticker", "date"],
)
panel_test_df = pd.read_parquet(
bench_dir / "panel_test.parquet", columns=["ticker", "date"],
)
panel_train_df["date"] = pd.to_datetime(panel_train_df["date"])
panel_test_df["date"] = pd.to_datetime(panel_test_df["date"])
split_date = panel_train_df["date"].max()
gt["event_date"] = pd.to_datetime(gt["event_date"])
# Restrict the eval/train pools to events whose ticker has at least
# ``min_history`` panel days BEFORE the event date in the combined
# panel. Without this filter, ~6.5% of train events sampled at the
# canonical step cannot produce a valid 63-day lookback at load
# time and were silently zero-padded then dropped.
min_history = max(_cfg_get_lookback(granularity), 63)
panel_full = pd.concat([panel_train_df, panel_test_df], ignore_index=True)
panel_full = panel_full.drop_duplicates(subset=["ticker", "date"])
panel_dates_per_ticker = (
panel_full.sort_values(["ticker", "date"]).groupby("ticker")["date"]
)
first_panel_date = panel_dates_per_ticker.first().to_dict()
def _has_lookback_history(row) -> bool:
first = first_panel_date.get(row["ticker"])
if first is None:
return False
# need at least min_history trading-day rows prior (use calendar
# days as a fast upper bound: 252 trading days ~ 365 calendar days).
return (row["event_date"] - first).days >= int(min_history * 1.45)
gt = gt[gt.apply(_has_lookback_history, axis=1)].copy()
train_pool = gt[gt["event_date"] <= split_date].copy()
eval_pool = gt[gt["event_date"] > split_date].copy()
# Sector and mcap_q come from the panel (across full date range).
full_panel = pd.read_parquet(
bench_dir / "panel_train.parquet",
columns=["ticker", "sector", "derived_market_cap"],
)
latest = full_panel.groupby("ticker").tail(1)[["ticker", "sector", "derived_market_cap"]]
latest["mcap_q"] = pd.qcut(
latest["derived_market_cap"].clip(lower=1),
4, labels=["Q1", "Q2", "Q3", "Q4"], duplicates="drop",
)
sector_map = dict(zip(latest["ticker"], latest["sector"]))
mcap_map = dict(zip(latest["ticker"], latest["mcap_q"]))
out: dict[Split, pd.DataFrame] = {}
for split, pool in (("train", train_pool), ("eval", eval_pool)):
pool = pool.copy()
pool["sector"] = pool["ticker"].map(sector_map)
pool["mcap_q"] = pool["ticker"].map(mcap_map)
n = n_train if split == "train" else n_eval
sampled = _stratified_sample(
pool, n, ["sector", "mcap_q", "event_type"], seed,
)
out[split] = sampled[["scenario_id", "ticker", "event_type"]]
return out
def _gen_t7(
granularity: str,
n_eval: int,
n_train: int,
seed: int,
) -> dict[Split, pd.DataFrame]:
"""T7 RE-Val: stratified by property_type x state."""
bench_dir = config.get_benchmark_dir(granularity)
train = pd.read_parquet(bench_dir / "re_train_properties.parquet")
eval_in = pd.read_parquet(bench_dir / "re_eval_inputs.parquet")
# ``re_train_properties`` carries 854 duplicate-address rows from
# multiple RentCast variants of the same listing. Dedup BEFORE
# sampling so the canonical eval set has unique addresses (the
# loader otherwise dedups, leaving the canon n short of budget).
if "address" in train.columns:
train = train.drop_duplicates(subset="address", keep="first").reset_index(drop=True)
if "address" in eval_in.columns:
eval_in = eval_in.drop_duplicates(subset="address", keep="first").reset_index(drop=True)
def _sample(df: pd.DataFrame, n: int) -> pd.DataFrame:
ptype_col = next(
(c for c in ("property_type", "propertyType", "type") if c in df.columns),
None,
)
state_col = next(
(c for c in ("state", "State") if c in df.columns), None,
)
addr_col = next(
(c for c in ("address", "addressLine1", "Address") if c in df.columns),
None,
)
strata = [c for c in (ptype_col, state_col) if c is not None]
if not strata or addr_col is None:
return df.sample(n=min(n, len(df)), random_state=seed).reset_index(drop=True)
sampled = _stratified_sample(df, n, strata, seed)
cols_to_keep = [addr_col] + strata
return sampled[cols_to_keep].rename(columns={addr_col: "address"})
return {"train": _sample(train, n_train), "eval": _sample(eval_in, n_eval)}
_GENERATORS = {
"T1": _gen_t1,
"T2": lambda g, ne, nt, s: _gen_t2_t5(g, "T2", ne, nt, s),
"T3": lambda g, ne, nt, s: _gen_t3_t6(g, "T3", ne, nt, s),
"T4": _gen_t4,
"T5": lambda g, ne, nt, s: _gen_t2_t5(g, "T5", ne, nt, s),
"T6": lambda g, ne, nt, s: _gen_t3_t6(g, "T6", ne, nt, s),
"T7": _gen_t7,
}
# ── Public API ────────────────────────────────────────────────────────────
def get_canonical_indices(
task: Task,
split: Split = "eval",
*,
granularity: str = "daily",
n_eval: dict[Task, int] | None = None,
n_train: dict[Task, int] | None = None,
seed: int | None = None,
) -> pd.DataFrame:
"""Return the canonical index list for `(task, split)`.
Reads from cache if available; otherwise generates, persists to cache,
and returns. The cache key encodes (n_eval, n_train, seed,
stratifier_version) so non-canonical re-tunes get their own cache dir.
Smoke-mode override: when ``MACROLENS_N_TRAIN`` and / or
``MACROLENS_N_EVAL`` env vars are set (positive int), they replace the
default budget for every task in this call. This lets the runner do an
end-to-end smoke (e.g. n_train=2, n_eval=1 across all 22 methods × 7
tasks) without touching the canonical cache or the CLI signature.
Explicit ``n_eval`` / ``n_train`` kwargs still take precedence.
"""
import os as _os
env_n_eval = _os.environ.get("MACROLENS_N_EVAL")
env_n_train = _os.environ.get("MACROLENS_N_TRAIN")
if n_eval is None and env_n_eval is not None:
try:
v = int(env_n_eval)
if v > 0:
n_eval = {t: v for t in EVAL_N_PER_TASK}
except ValueError:
pass
if n_train is None and env_n_train is not None:
try:
v = int(env_n_train)
if v > 0:
n_train = {t: v for t in TRAIN_N_PER_TASK}
except ValueError:
pass
n_eval_map = n_eval or EVAL_N_PER_TASK
n_train_map = n_train or TRAIN_N_PER_TASK
s = SEED if seed is None else seed
key = budgets.cache_key(n_eval=n_eval_map, n_train=n_train_map, seed=s)
cache_dir = _cache_dir(granularity, key)
cache_path = cache_dir / f"{split}_{task}.parquet"
if cache_path.exists():
return pd.read_parquet(cache_path)
# Cache miss: generate both splits for this task and persist.
gen = _GENERATORS[task]
pair = gen(granularity, n_eval_map[task], n_train_map[task], s)
cache_dir.mkdir(parents=True, exist_ok=True)
for sp, df in pair.items():
df.to_parquet(cache_dir / f"{sp}_{task}.parquet", index=False)
logger.info(
"Canonical-indices cache write: %s (%d rows)",
cache_dir / f"{sp}_{task}.parquet", len(df),
)
return pair[split]
def build_all(
granularity: str = "daily",
*,
n_eval: dict[Task, int] | None = None,
n_train: dict[Task, int] | None = None,
seed: int | None = None,
) -> dict[str, int]:
"""Build canonical indices for every (task, split) pair.
Returns a summary dict mapping `<task>_<split>` -> n_rows. Idempotent:
re-running with the same budgets is a no-op (cache hits).
"""
summary: dict[str, int] = {}
for task in ("T1", "T2", "T3", "T4", "T5", "T6", "T7"):
for split in ("train", "eval"):
df = get_canonical_indices(
task, split, granularity=granularity,
n_eval=n_eval, n_train=n_train, seed=seed,
)
summary[f"{task}_{split}"] = len(df)
return summary