File size: 12,048 Bytes
3f316da | 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 | #!/usr/bin/env python3
"""Summarize source-separated FI-2010 and LOBSTER prediction evidence."""
from __future__ import annotations
import argparse
import csv
import json
from pathlib import Path
import numpy as np
METHODS = ("timae", "lobgrid_timae", "simlob2026")
METHOD_LABELS = {
"timae": "Ti-MAE-style",
"lobgrid_timae": "LOBGrid--Ti-MAE-style",
"simlob2026": "SimLOB-style",
}
SEEDS = (2026, 2027, 2028)
SETTINGS = ("from_scratch", "fine_tune")
CLASS_FIELDS = tuple(
f"{metric}_{name}"
for metric in ("precision", "recall", "f1")
for name in ("down", "stationary", "up")
)
def load_results(directory: Path) -> list[dict]:
rows = []
for path in sorted(directory.glob("*.json")):
if path.name.endswith(".pretrain.json"):
continue
payload = json.loads(path.read_text(encoding="utf-8"))
if payload.get("setting") in SETTINGS and "test" in payload:
payload["_path"] = str(path)
rows.append(payload)
return rows
def validate(rows: list[dict], assets: int, source: str) -> None:
expected = assets * len(METHODS) * len(SEEDS) * len(SETTINGS)
keys = [(r["stock"], r["method"], r["seed"], r["setting"]) for r in rows]
if len(keys) != expected or len(set(keys)) != expected:
raise ValueError(f"{source}: found {len(keys)} unique cases; expected {expected}")
if {r["method"] for r in rows} != set(METHODS):
raise ValueError(f"{source}: method set is incomplete")
if {int(r["seed"]) for r in rows} != set(SEEDS):
raise ValueError(f"{source}: seed set is incomplete")
def bootstrap(values: list[float], seed: int = 20260801) -> list[float]:
array = np.asarray(values, dtype=np.float64)
rng = np.random.default_rng(seed)
draws = rng.choice(array, size=(10000, len(array)), replace=True).mean(axis=1)
return [float(x) for x in np.quantile(draws, [0.025, 0.975])]
def condition_metrics(rows: list[dict]) -> dict:
result = {
"macro_f1_mean": float(np.mean([r["test"]["macro_f1"] for r in rows])),
"macro_f1_sd": float(np.std([r["test"]["macro_f1"] for r in rows], ddof=1)),
"balanced_accuracy_mean": float(
np.mean([r["test"]["balanced_accuracy"] for r in rows])
),
"stopped_epoch_mean": float(np.mean([r.get("stopped_epoch", 0) for r in rows])),
"optimizer_steps_mean": float(np.mean([r.get("optimizer_steps", 0) for r in rows])),
}
for field in CLASS_FIELDS:
result[field] = (
float(np.mean([r["test"][field] for r in rows]))
if all(field in r["test"] for r in rows)
else None
)
result["pooled_confusion_rows_true_cols_pred"] = (
np.asarray(
[r["test_confusion"]["matrix_rows_true_cols_pred"] for r in rows],
dtype=np.int64,
).sum(axis=0).tolist()
if all("test_confusion" in r for r in rows)
else None
)
return result
def paired_source_summary(rows: list[dict], source: str) -> dict:
index = {
(r["stock"], r["method"], int(r["seed"]), r["setting"]): r
for r in rows
}
assets = sorted({r["stock"] for r in rows})
output = {}
for method in METHODS:
asset_seed = {}
for asset in assets:
for seed in SEEDS:
scratch = index[(asset, method, seed, "from_scratch")]
tuned = index[(asset, method, seed, "fine_tune")]
asset_seed[(asset, seed)] = float(
tuned["test"]["macro_f1"] - scratch["test"]["macro_f1"]
)
asset_means = {
asset: float(np.mean([asset_seed[(asset, seed)] for seed in SEEDS]))
for asset in assets
}
unit_values = list(asset_means.values()) if len(assets) > 1 else list(asset_seed.values())
day_effects = {}
sample = next(
(r for r in rows if r["method"] == method and r.get("test_by_group")),
None,
)
for day in sorted(sample["test_by_group"] if sample else {}):
values = []
for asset in assets:
for seed in SEEDS:
scratch = index[(asset, method, seed, "from_scratch")]
tuned = index[(asset, method, seed, "fine_tune")]
values.append(
float(
tuned["test_by_group"][day]["metrics"]["macro_f1"]
- scratch["test_by_group"][day]["metrics"]["macro_f1"]
)
)
day_effects[day] = {
"mean": float(np.mean(values)),
"sd": float(np.std(values, ddof=1)),
}
method_rows = [r for r in rows if r["method"] == method]
output[method] = {
"model": METHOD_LABELS[method],
"source": source,
"reporting_unit": "initialization seed" if len(assets) == 1 else "asset after seed averaging",
"mean_delta": float(np.mean(unit_values)),
"sd_delta": float(np.std(unit_values, ddof=1)),
"unit_resampling_interval_95": bootstrap(unit_values),
"wins": int(sum(x > 0 for x in unit_values)),
"units": len(unit_values),
"asset_mean_deltas": asset_means,
"asset_seed_deltas": {f"{a}|{s}": v for (a, s), v in asset_seed.items()},
"test_day_effects": day_effects,
"conditions": {
setting: condition_metrics([r for r in method_rows if r["setting"] == setting])
for setting in SETTINGS
},
}
return output
def write_csv(path: Path, rows: list[dict]) -> None:
fields = []
for row in rows:
for field in row:
if field not in fields:
fields.append(field)
with path.open("w", newline="", encoding="utf-8") as handle:
writer = csv.DictWriter(handle, fieldnames=fields)
writer.writeheader()
writer.writerows(rows)
def main() -> None:
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("--fi-dir", required=True)
parser.add_argument("--lobster-dir", required=True)
parser.add_argument("--fi-zscore-dir")
parser.add_argument("--output-dir", required=True)
args = parser.parse_args()
fi_rows = load_results(Path(args.fi_dir))
lobster_rows = load_results(Path(args.lobster_dir))
validate(fi_rows, assets=1, source="FI-2010 DecPre train-only")
validate(lobster_rows, assets=4, source="LOBSTER")
payload = {
"FI-2010 DecPre train-only": paired_source_summary(
fi_rows, "FI-2010 DecPre train-only"
),
"LOBSTER": paired_source_summary(lobster_rows, "LOBSTER"),
}
if args.fi_zscore_dir:
zscore_rows = load_results(Path(args.fi_zscore_dir))
validate(zscore_rows, assets=1, source="FI-2010 official Zscore sensitivity")
payload["FI-2010 official Zscore sensitivity"] = paired_source_summary(
zscore_rows, "FI-2010 official Zscore sensitivity"
)
output_dir = Path(args.output_dir)
output_dir.mkdir(parents=True, exist_ok=True)
(output_dir / "external_source_evidence.json").write_text(
json.dumps(payload, indent=2), encoding="utf-8"
)
flat_rows = []
for source, methods in payload.items():
for method, summary in methods.items():
scratch = summary["conditions"]["from_scratch"]
tuned = summary["conditions"]["fine_tune"]
flat_rows.append(
{
"source": source,
"method": method,
"model": summary["model"],
"scratch_macro_f1": scratch["macro_f1_mean"],
"target_local_macro_f1": tuned["macro_f1_mean"],
"paired_delta": summary["mean_delta"],
"paired_delta_sd": summary["sd_delta"],
"ci_low": summary["unit_resampling_interval_95"][0],
"ci_high": summary["unit_resampling_interval_95"][1],
"wins": summary["wins"],
"units": summary["units"],
}
)
write_csv(output_dir / "external_source_evidence.csv", flat_rows)
paired_case_rows = []
for source, rows in (
("FI-2010 DecPre train-only", fi_rows),
("LOBSTER", lobster_rows),
):
index = {
(row["stock"], row["method"], int(row["seed"]), row["setting"]): row
for row in rows
}
for stock in sorted({row["stock"] for row in rows}):
for method in METHODS:
for seed in SEEDS:
scratch = index[(stock, method, seed, "from_scratch")]
tuned = index[(stock, method, seed, "fine_tune")]
paired_case_rows.append(
{
"source": source,
"unit": stock,
"method": method,
"seed": seed,
"scratch_macro_f1": scratch["test"]["macro_f1"],
"target_local_macro_f1": tuned["test"]["macro_f1"],
"paired_delta": (
tuned["test"]["macro_f1"]
- scratch["test"]["macro_f1"]
),
"scratch_stopped_epoch": scratch.get("stopped_epoch"),
"scratch_optimizer_steps": scratch.get("optimizer_steps"),
"target_local_stopped_epoch": tuned.get("stopped_epoch"),
"target_local_optimizer_steps": tuned.get("optimizer_steps"),
}
)
write_csv(output_dir / "external_paired_seed_cases.csv", paired_case_rows)
class_rows = []
day_rows = []
for source in ("FI-2010 DecPre train-only", "LOBSTER"):
for method, summary in payload[source].items():
for setting, condition in summary["conditions"].items():
class_rows.append(
{
"source": source,
"method": method,
"setting": setting,
**{field: condition[field] for field in CLASS_FIELDS},
"confusion_rows_true_cols_pred": json.dumps(
condition["pooled_confusion_rows_true_cols_pred"]
),
"mean_stopped_epoch": condition["stopped_epoch_mean"],
"mean_optimizer_steps": condition["optimizer_steps_mean"],
}
)
for day, day_summary in summary["test_day_effects"].items():
day_rows.append(
{
"source": source,
"method": method,
"test_day": day,
"mean_delta": day_summary["mean"],
"sd_delta": day_summary["sd"],
}
)
write_csv(output_dir / "external_class_and_stopping_metrics.csv", class_rows)
write_csv(output_dir / "external_test_day_effects.csv", day_rows)
lines = [
"# Source-separated external evidence",
"",
"| source | model | scratch F1 | local F1 | paired delta ± SD | wins |",
"|---|---|---:|---:|---:|---:|",
]
for row in flat_rows:
lines.append(
f"| {row['source']} | {row['model']} | {row['scratch_macro_f1']:.4f} "
f"| {row['target_local_macro_f1']:.4f} | {row['paired_delta']:+.4f}"
f"±{row['paired_delta_sd']:.4f} | {row['wins']}/{row['units']} |"
)
(output_dir / "external_source_evidence.md").write_text(
"\n".join(lines) + "\n", encoding="utf-8"
)
if __name__ == "__main__":
main()
|