Anyone can experiment with a lightweight yet relatively robust architecture of TCN-Attention for financial market prediction (primarily trend direction and next-bar price estimation). we can sharpen the setup on potential overfitting, validation strategies, and feature engineering.
Dataset:
60,000+ rows of 1h OHLCV data (any scrip, but the framework is generic).
Train / val / test split: 70/15/15 (chronological, no shuffle).
Architecture can be:
Input (lookback=128)
→ Conv1D (filters=64, kernel=8, dilation=1)
→ TCN block (dilations=1,2,4,8,16, residual)
→ Multi-head Attention (query from last timestep, key/value from full sequence)
→ Global Avg Pooling (or flatten)
→ Dense(32, dropout=0.3)
→ Output:
- Regression: price change % (tanh scaled)
- Classification: direction (up/down, sigmoid)
Loss: combined MSE + binary crossentropy (weighted).
Results that are achievable:
Directional accuracy: ~54–56% (varies by pair/period, not great but above coin flip).
Regression MAE (price change %): ~0.32% for crypto, 0.09% for forex.
Sharpe of strategy based on model signals (backtest): ~1.1 (after 0.1% slippage).
Go and google similar ML models...