Closing Auction Prediction

Research project predicting Nasdaq closing-cross price moves from pre-close order-imbalance data, and backtesting a long/short strategy built on those predictions.

Data

data/202503_imbalance.tar — Nasdaq closing-auction imbalance feed, 21 trading days (2025-03-03 to 2025-03-31), snapshots every 10s from 15:50-15:55 ET and every 1s from 15:55 to 16:00 ET. 500 symbols/day (562 unique across the month, universe rotates daily), ~4.1M rows raw.

Target: target_bps_vs_mid = (close_price / mid - 1) * 1e4, where mid = (bid + ask) / 2 as of the last pre-close snapshot. Idealized and cost-free by construction — chosen over raw close_price because price levels aren't comparable across symbols, and over targets referenced to Nasdaq's own auction-price estimates (near_price/far_price) because that's a different question from predicting the move itself.

Final feature set (selected in 03) via lasso bootstrap t-stats + SHAP/SAGE on a LightGBM model): near_price_vs_mid_bps, urgency, ref_price_vs_mid_bps, far_price_vs_mid_bps. Full feature glossary and selection rationale in 02)'s Feature Reference and 03)'s Conclusion.

Result (see 06)): an MLP + quintile long/short bucket strategy entered at 15:58:00 posts Sharpe 14.5 (annualized) on the 17-day train period and a directionally consistent but not statistically significant Sharpe 5.6 on the 4-day held-out test period (p = 0.53). A promising research result, not yet a validated trading strategy — execution feasibility in particular is still open (NASDAQ's MOC/LOC order cutoff is 15:55:00, three minutes before this entry).

Workflow

Notebooks run in numbered order, each building on the last:

1. 01) data_exploration.ipynb — data inventory, univariate analysis, data quality issues, lookahead-bias check (cross is the auction print itself, delivered in-band pre-close as NaN — never forward-fill it), and panel construction. Findings written up in docs/1) data_exploration_report.md.
2. 02) features_and_target.ipynb — feature engineering (imbalance/spread/urgency, ADV-normalized sizes, short-horizon momentum and rolling vol), target definition, train/test split, correlation analysis, pruning of non-stationary/scale-variant raw price levels.
3. 03) models_and_feature_selection.ipynb — expanding-window CV, elastic net with bootstrap t-stats, a small LightGBM model with SHAP/SAGE, cross-checked against each other to pick the final 4-feature set, then a LazyPredict scan to shortlist model families.
4. 04) train_wandb.ipynb — Bayesian hyperparameter sweeps (Weights & Biases, 100 trials each) over the shortlisted families: Huber, Lasso, GradientBoosting, RandomForest, and a PyTorch MLP. Best configs per family saved to best_hyperparameters.json.
5. 05) backtest.ipynb — long/short bucket backtest using the tuned models: realized P&L (long fills at ask, short at bid, both exit via a guaranteed-fill Market-on-Close order at close_price, not the idealized mid-based training target), walk-forward folds, swept across 4 bucket-construction schemes (50/50, quartile, quintile, decile) × 5 tuned model classes × 6 entry timings (15:55:00 through 15:59:00 one minute apart, plus the terminal 15:59:59) to check whether performance survives trading earlier than the last second before close. Adds an execution-latency check (repricing an already-selected bucket 1/5/10s after the signal) to confirm the chosen entry timing isn't a zero-latency artifact. Lands on MLP + quintile buckets + 15:58:00 entry as the final recipe.
6. 06) test.ipynb — final validation of that recipe: identical backtest run on both the 17-day train period (walk-forward retrained) and the 4-day held-out test period (single model frozen after training once on all of train), with full P&L/Sharpe/t-stat/hit-rate stats plus stress tests (removing vs. winsorizing top winners, execution lag, bootstrap Sharpe CI, max drawdown, sub-period stability). Also where the MLP training loop's eval-fold leakage — present in 05)'s copy — gets fixed, so this notebook's numbers are the ones to trust.

Repository layout

01) data_exploration.ipynb          EDA, data quality, lookahead check
02) features_and_target.ipynb       Feature engineering, target, train/test split
03) models_and_feature_selection.ipynb   CV, elastic net, SHAP/SAGE, final feature set
04) train_wandb.ipynb               W&B hyperparameter sweeps per model family
05) backtest.ipynb                  Long/short backtest, walk-forward, entry-timing +
                                     execution-lag robustness, picks the final recipe
06) test.ipynb                      Final recipe validated on train + held-out test
best_hyperparameters.json           Best config per model family, from 04)
data/                               Raw + processed data (mostly gitignored, see below)
docs/                               Take-home brief + 01)'s write-up (data_exploration_report.md)
research/                           Background papers, gitignored — see research/README.md
requirements.txt

data/ is gitignored except panel_clean.csv, train.csv, test.csv, and columns.json — the raw .tar and any other derived files are excluded. research/*.pdf is gitignored too; research/README.md has full citations and links to what's there.

Setup

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

04) logs sweeps to Weights & Biases — run wandb login once before executing that notebook.

Working notes

• No production concerns here — this is research/EDA code, optimized for correctness of conclusions over robustness to edge cases.
• Lookahead bias is treated as the primary risk throughout (see 01) Section 4 in particular); each notebook that touches CV or the target documents how it avoids leaking future information into a fold.