Dual-model AI system with dedicated LONG and SHORT neural networks, achieving 84.9% walk-forward validated accuracy. Each direction has its own specialized model trained on distinct market patterns. No overfitting. No cherry-picking. Just rigorous machine learning.
Five battle-tested stages transform raw market data into actionable trading signals, every hour, 24/7.
12 months of 1-hour candles from 25 major cryptocurrencies. OHLCV data sourced directly from exchange APIs with strict quality checks for gaps, outliers, and exchange downtime.
~216,000 candles across 25 coinsUp to 23 carefully engineered market indicators including momentum, volatility, volume patterns, crash velocity, and sell pressure signals. Each feature is normalized and statistically validated to avoid data leakage.
23 indicators • zero lookahead biasBidirectional LSTM with multi-head attention mechanism, trained on GPU-accelerated infrastructure. The attention layer learns which time-steps and features matter most for each prediction.
BiLSTM + Attention • GPU accelerated4-window temporal validation ensuring the model never sees future data. Each window trains on historical data and tests on the subsequent unseen period. No cherry-picking, no look-ahead bias.
4 windows • temporal integrityXGBoost + LightGBM + dedicated LONG LSTM + dedicated SHORT LSTM combined for robust predictions. Each direction uses a specialized neural network: the LONG model excels at identifying uptrends, while the SHORT model is trained specifically on crash patterns with additional sell-pressure features.
4 models • direction-specializedWhy one model isn't enough — and how dedicated LONG and SHORT networks outperform generic approaches.
Specialized in identifying uptrend setups: momentum continuation, breakout confirmation, and mean-reversion opportunities.
Specialized in detecting crash setups: panic selling, volume spikes on red candles, cascading liquidations, and bearish structure breakdowns.
Crypto markets are inherently asymmetric: prices tend to rise gradually during uptrends but crash violently during sell-offs. A single model trained on both directions learns a "compromise" that excels at neither. Our dual-model approach trains each network on the specific patterns it needs to recognize — the LONG model learns momentum and accumulation signals, while the SHORT model learns panic-selling signatures, volume anomalies, and structural breakdowns that precede drops. The result: higher precision for both directions, fewer false signals, and better risk management.
Why most backtests lie, and how we ensure our results are real.
Traditional backtesting trains and tests on the same dataset — or uses random splits that leak future information into the past. This produces inflated accuracy numbers that fall apart in live trading.
Walk-forward validation mimics real-world conditions: the model only ever trains on past data and is tested on data it has never seen. We expand the training window over time and test on the next period, exactly as the model would operate live.
A high-level view of the data flow from raw candles to actionable prediction.
Simplified architecture diagram. Internal hyperparameters and layer configurations are proprietary.
Accuracy measured on truly unseen data across four temporal windows.
| Window | Training Data | Accuracy Range | Visual |
|---|---|---|---|
| Window 1 | 3 months | 68 – 72% | |
| Window 2 | 6 months | 71 – 73% | |
| Window 3 | 9 months | 80 – 83% | |
| Window 4 | 12 months | 81 – 84% |
Rule-based bots execute YOUR rules. DeepAlpha PREDICTS price direction using deep learning.
| Feature | 3Commas / Cryptohopper | DeepAlpha |
|---|---|---|
| Intelligence | Rule-based (if/then) | ✓ Deep Learning AI |
| Prediction | ✗ None — executes your rules | ✓ Predicts price direction |
| Validation | ✗ Simple backtest | ✓ Walk-forward (4 windows) |
| Adapts to Market | ✗ Static rules | ✓ Weekly retraining |
| Ensemble | ✗ Single strategy | ✓ XGBoost + LGBM + LSTM |
| Open Source | ✗ Closed source | ✓ Core on GitHub |
| Custody | API keys on their servers | ✓ Non-custodial |
| Price | $29 – $99/mo | ✓ Free trial, then $39/mo |
Markets change. Our AI evolves with them.
Fresh market data is ingested every week. The model retrains on the latest price action, volume patterns, and market regimes to stay current.
Training runs on dedicated GPU infrastructure, allowing us to iterate rapidly on architectures, hyperparameters, and feature sets without bottlenecks.
New features and architectures are tested monthly. From Transformer variants to GNN-based cross-asset modeling, we continuously push the boundaries of what's possible.
See DeepAlpha's predictions in action. No credit card required.
No credit card • Cancel anytime • Non-custodial