Audit of Ml and Econometric Models on Financial Time Series

A quantitative research platform trains 24 model families every night on daily market data, walk-forward, across about 1,000 assets. We are looking for an independent expert to check that these models are implemented correctly and evaluated honestly.

The work

• ML models: XGBoost, LightGBM, Random Forest, Extra Trees, regularized linear models, SVR, MLP, LSTM, TCN.
• Econometric models: GARCH, EGARCH, GJR-GARCH, SARIMAX/ARIMA, VAR, VECM, Kalman.
• What to check: does each model do what its configuration says? Are the hyperparameter search spaces (Optuna) suitable for financial targets, and is the search free of look-ahead?
• Model collapse: find where a fit degenerates into a constant predictor (the mean, stump-only trees, zero coefficients, the maximum-regularization corner, a failed GARCH fit recorded as success), and prove it on synthetic data.
• Targets: forward returns and forward realized volatility at 1, 5, 21, 63, 126 and 252 days (regression), plus a design review of an upcoming classification pipeline (calibration, class imbalance, AUC, Brier score).
• Evaluation: how to correct the metrics when daily predictions cover overlapping horizons (Newey–West / Hansen–Hodrick, purged and embargoed cross-validation), and which benchmarks and loss functions to use for volatility (HAR-RV, QLIKE, Diebold–Mariano).

Deliverables

• Phase A, fixed price: a prioritized report with file and line references, a collapse-risk matrix, an evaluation-method note and a classification design note.
• Phase B, time and materials: fixes as small pull requests, each with a test.