Ml4t Model Validation

Model Validation Workflow

A model that passes a single train/test split proves nothing. Rigorous validation requires combinatorial CV, overfitting probability, deflated statistics, feature attribution, and out-of-time holdout - all before any backtest.

The Problem

A researcher splits data 80/20, trains a model, sees good test-set performance, and runs a backtest. The backtest looks promising. They deploy. The strategy loses money immediately. The cause: the single split was lucky, the model memorized regime-specific patterns, and hyperparameter tuning leaked information across the boundary. Without multiple validation gates, a model that looks good on one split can be arbitrarily overfit.

The Pattern

WRONG

# Single train/test split, no overfitting checks, straight to deployment
from sklearn.model_selection import train_test_split
import lightgbm as lgb
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, shuffle=True)
model = lgb.LGBMRegressor().fit(X_train, y_train)
score = model.score(X_test, y_test)
print(f"R2: {score:.3f}")  # 0.15 - good enough, deploy

CORRECT

import numpy as np
import lightgbm as lgb
from scipy.stats import norm

# Gate 1: CPCV - multiple train/test paths, not one split (see ml4t-cpcv)
cv_sharpes = []
for train_idx, test_idx in time_aware_cv_splits:  # C(10,2) = 45 splits
    model = lgb.LGBMRegressor(n_estimators=100, random_state=42)
    model.fit(X[train_idx], y[train_idx])
    preds = model.predict(X[test_idx])
    sharpe = np.mean(preds * y[test_idx]) / np.std(preds * y[test_idx]) * np.sqrt(252)
    cv_sharpes.append(sharpe)

assert np.median(cv_sharpes) > 0, "median path Sharpe is not positive"  # Gate 1

# Gate 2: fraction of paths that lose money. This is NOT PBO, which ranks the
# in-sample winner out of sample across splits (see ml4t-backtest-overfitting).
loss_rate = np.mean([s < 0 for s in cv_sharpes])
assert loss_rate < 0.50, f"{loss_rate:.0%} of paths negative - likely overfit"

# Gate 3: the bound applies to the CONFIGURATIONS you chose between, not to
# CPCV paths of one model - those are correlated estimates of the same number.
trial_sharpes = [np.median(paths) for paths in cv_sharpes_per_config]  # ALL tried
n = len(trial_sharpes)
assert n > 1, "a selection bound needs more than one trial"
expected_max = np.std(trial_sharpes) * (
    (1 - np.euler_gamma) * norm.ppf(1 - 1 / n)
    + np.euler_gamma * norm.ppf(1 - 1 / (n * np.e))
)
best = int(np.argmax(trial_sharpes))
assert trial_sharpes[best] > expected_max, "best config is inside the bound"

# Refit the SELECTED configuration; `model` is just the last CV fold's leftover
final = lgb.LGBMRegressor(**configs[best]).fit(X, y)

# Gate 4: SHAP - verify features match hypothesis (see ml4t-shap-analysis)
import shap
shap_values = shap.TreeExplainer(final).shap_values(X)

# Gate 5: OOS holdout - data never seen in any CV fold
oos_sharpe = (np.mean(final.predict(X_holdout) * y_holdout)
              / np.std(final.predict(X_holdout) * y_holdout) * np.sqrt(252))
degradation = (np.mean(cv_sharpes) - oos_sharpe) / np.mean(cv_sharpes)
assert degradation < 0.30, f"OOS degradation {degradation:.0%} - too high"

Gate Summary

Gates are sequential. Do not skip to Gate 5 hoping a good holdout compensates for Gate 2.

Guardrails

• If SHAP shows the model relies on a single feature for > 40% of predictions, the model is fragile
• If OOS degradation is < 5%, be suspicious - it often means data leakage, not a great model
• If CV Sharpe variance across folds is > 1.0, the signal is unstable across regimes

Production Implementation

ml4t-diagnostic provides CPCV splitting with fold Sharpes and DSR:

from ml4t.diagnostic.api import ValidatedCrossValidation
from ml4t.diagnostic.config import ValidatedCrossValidationConfig

config = ValidatedCrossValidationConfig(n_groups=10, n_test_groups=2, embargo_pct=0.01)
vcv = ValidatedCrossValidation(config)
result = vcv.fit_evaluate(X, y, model, times=timestamps)
fold_sharpes = [fold.sharpe_ratio for fold in result.fold_results]

Checklist

• Cross-validation uses CPCV with purging and embargo, not random splits
• Loss rate across CPCV paths < 50% (PBO itself: ml4t-backtest-overfitting)
• Best configuration clears the selection bound for the number tried
• SHAP feature importance aligns with economic hypothesis
• True out-of-time holdout tested (data never used in any CV fold)
• OOS performance degradation < 30% from in-sample