3-class profit-maximizing gold ML pipeline (triple-barrier, $100k account, feature search)

[ataridan7-cpu]

Overview

Rewrites gold_direction_ml.ipynb into a both-directions, profit-maximizing pipeline that trades a $100,000 toy account where the model decides direction and size.

Architecture

Labels — 3-class triple-barrier (every bar)

Every bar is an event (no directional pre-filter). Each entry gets two volatility-scaled barriers and a time barrier; the first touched determines the label:

• +1 (up) — upper barrier entry·(1 + pt_mult·σ) hit first → go long
• −1 (down) — lower barrier entry·(1 − sl_mult·σ) hit first → go short
• 0 (flat) — time barrier (max_holding_days) → stay flat

Model — multiclass XGBoost (multi:softprob, 3 classes)

Outputs [P(down), P(flat), P(up)]. Directional edge = P(up) − P(down).

Position sizing — confidence-proportional, compounded

frac = min(1.0, size_scale · |edge|) of current equity (no leverage). Long on edge > tau, short on edge < −tau, flat otherwise. Round-trip costs deducted per trade.

Optimization — pure profit (final account value)

Optuna, the tau/scale grid, and the feature-combo search all maximize final $ equity — no Sharpe penalty.

Feature search — exhaustive (127 subsets) + greedy forward

Searches all non-empty subsets of 7 candidates by OOF profit with quick fixed params. Greedy forward is run for comparison; the exhaustive winner is locked in when they disagree.

Purged walk-forward CV

Expanding folds with purging + embargo (de Prado-style), preventing label-overlap lookahead.

Model comparison

XGBoost vs LightGBM vs multinomial logistic vs LSTM, all scored by $ profit on the untouched holdout. SHAP summary for the XGBoost up-class.

$100k backtest

Daily equity curve, Sharpe / Sortino / max-DD / Calmar, trade stats (hit rate, avg win/loss, profit factor, long vs short counts), and per-regime breakdown (vol regime × trend).

Bug fixes applied (from high-effort code review)

#	Issue	Fix
1	Sample weights computed on full dataset before split — holdout event windows inflated concurrency counts for late training bars	Weights now computed on events.iloc[:split] only
2	Final model trained without early stopping using Optuna's suggested n_estimators ceiling — overtrained vs CV fold models	Final fit uses fit_with_es (same early-stopping logic as CV folds)
3	Logistic baseline received no sample_weight — unfair comparison	Passes logisticregression__sample_weight=w_tr via pipeline kwargs
4	simulate_from_edges silently accepted entry_times outside events.index, setting last_exit=NaT and disabling the non-overlap guard	Raises ValueError immediately on any NaT in the reindexed result
5	Empty trades DataFrame → groupby on zero rows returned silent empty table	Guard added with explicit "no trades taken" message
6	subset_profit referenced X_tr which gets narrowed later — partial notebook reruns KeyErrored on the feature search	_X_search = X_tr snapshot taken before narrowing
7	t1pos_tr clipped to split−1, causing over-purging of late-training events in early CV folds	Clip removed; true label-end positions flow through
8	Greedy search called subset_profit(winner) twice — second call wasted a full OOF CV run	Single max(generator) captures both winner and value
9	Sharpe/max-DD computed from uniform-spread equity curve with no caveat	Comment added: approximation understates intra-trade drawdowns
10	Regime breakdown accessed Trend_Sign with no empty-trades guard	Covered by fix jonathan-bower#5

Runtime

~53 s with n_trials=20, n_splits=5, exhaustive feature search over 127 subsets.

https://claude.ai/code/session_01SbRD1crgrw43xnCTpzRWZr