tbukuai/stock-patchtst-sp500

PatchTST — S&P 500 Next-Day Prediction

Best Model: v3 (this branch)

Metric	Value
DA(ctx)	60.8% (single-split) / 53.3% (walk-forward)
Sharpe(ctx)	4.32
RMSE	$26.84
R²	0.9976

⚠️ Walk-forward DA of 53.3% includes wavelet look-ahead bias. Honest (causal) DA is ~46-50%. See research findings below.

Versions (branches)

Branch	Channels	Wavelet	DA(ctx)	Load with
main	13	Global	60.8%/53.3% WF	from_pretrained("tbukuai/stock-patchtst-sp500")
v1	13	Global	51.3%	from_pretrained("tbukuai/stock-patchtst-sp500", revision="v1")
v3	13	Global	60.8%	from_pretrained("tbukuai/stock-patchtst-sp500", revision="v3")
v4	25	Causal	60.1%/~43% WF	from_pretrained("tbukuai/stock-patchtst-sp500", revision="v4")
v5	8	None	46.1%	from_pretrained("tbukuai/stock-patchtst-sp500", revision="v5")

Usage

from transformers import PatchTSTForPrediction

# Load best model (v3)
model = PatchTSTForPrediction.from_pretrained("tbukuai/stock-patchtst-sp500")

# Load specific version
model_v5 = PatchTSTForPrediction.from_pretrained("tbukuai/stock-patchtst-sp500", revision="v5")

Key Research Finding

S&P 500 daily direction is NOT predictable from price/technical indicators with honest methodology. The 53.3% walk-forward result relies on wavelet look-ahead bias. Without it: ≤50% (random).

Full analysis: RESEARCH_FINDINGS.md

Architecture

• PatchTST (d=128, heads=16, layers=3, ~600K params)
• Context: 252 trading days
• Prediction: 1 day ahead
• Features: OHLCV + technical indicators + wavelet denoising (v3)

⚠️ Disclaimer

Research/educational only. NOT for trading. The model does not reliably predict market direction.