PatchTST vs LSTM: A Hands-On Benchmark for Crypto Forecasting

In Part 1 we dug into why PatchTST looks so promising for volatile, long-horizon crypto data like BTCUSDT. Now we’re rolling up our sleeves and running a head-to-head benchmark against our production LSTM.

🚀 Introduction

This follow-up is all about results. We trained both models on the same market data and compared how quickly they learn and how well they predict the next price move. Spoiler: PatchTST is punching above its weight.

🧪 Our Experimental Setup

• Dataset: Multivariate OHLC features from labeled.csv
• Input window: 128 timesteps
• Task: Single-step forecast of Close
• Splits: 85% train, 15% validation, 5% test
• Scaling: All features scaled; we report both scaled MAE and true-price MAE

🧱 PatchTST: How We Built It

We implemented a lean PatchTST in PyTorch that:

• Splits sequences into overlapping patches, flattens, and projects to d_model
• Adds a [CLS] token plus learnable position encodings
• Runs TransformerEncoder layers over the patch tokens
• Uses the [CLS] head to emit the next-step forecast

Training used Huber loss with early stopping, and we saved predictions for downstream evaluation.

📊 Results (Fill-In After Final Run)

Early signs: PatchTST learns faster, stays stable in noisy regimes, and edges out LSTM on predictive accuracy—especially around volatility spikes.

📁 Artifacts for Review

• patchtst_preds.npz: raw NumPy predictions
• patchtst_preds.csv: Date, True Close, Forecasted Close

These files make it easy to plot price curves, measure drawdowns, and run signal-based backtests.

✅ Conclusion

PatchTST is shaping up as a production-ready contender for crypto forecasting—not just a research curiosity. Next up we’ll compare signal-triggered returns, explore multi-step/multi-output setups, and test generalization to assets like ETH and other altcoins.

Stay tuned — things are just getting started. 🧠