tbukuai/patchtst-wavelet-sp500-research

PatchTST + Wavelet S&P 500 Research

Final Conclusion (v1-v6): S&P 500 daily direction cannot be predicted from price/technical data.

Results Summary

Version	Features	Wavelet	DA(ctx) Single	DA(ctx) Walk-Forward	Verdict
v1	13ch OHLCV+tech	Global	51.3%	—	DA(diff)=69% was fake
v2	13ch	Global+MADL	55.3%	—	Collapsed to always-up
v3	13ch	Global	60.8%	53.3%	Look-ahead bias
v4	25ch+VIX	Causal	60.1%	~43.8%	More = worse
v5	8ch minimal	None	46.1%	—	Below random
v6	12ch custom	None	51.9%	49.6%	Random

v6 Features Tested

OHLCV + MA5 + MA23 + MA53 + RSI + MACD + VIX + MAVOL = 12 channels

Multi-scale moving averages, RSI, MACD, VIX, and volume smoothing — none break 50% DA.

What Actually Works

Cross-sectional ranking (not direction prediction):

• LightGBM on S&P 500: IC=0.02, Sharpe 0.47-1.07
• Signal is in slow factors (60-day volatility, momentum), not daily direction

Key Findings

1. DA(diff) ≈ 70% is fake — always use DA(ctx)
2. Global wavelet = look-ahead bias — v3's 53.3% was 100% from leakage
3. More features = overfitting — v4 (25ch) < v5 (8ch) < v6 (12ch) ≈ random
4. Custom technical indicators don't help — v6 = 49.6% WF = random
5. Simple > complex — LightGBM (2 sec) > PatchTST (hours) > Kronos (102M params)

Repo Contents

• notebooks/ — v3-v6 Colab training notebooks
• results/ — v3-v6 results JSON files
• PROJECT_CONCLUSION.md — Full findings + future work
• CLAUDE.md — Agent reference (lessons, pitfalls, code)
• train_v6.py — Standalone training script