Below we will share some of our technology details to prove, what we did and what we achieved so far.
A conceptual framework and a model were built that take into account the following metrics:
1. Required length of historical training data: 20000 samples, i.e.:
- 3 years for crypto 1h candles,
- 10 years for stocks market 1h candles.
2. Prediction accuracy/prediction error (in %):
- Max / Min price prediction: f1_score : 45-50% (enough to have profit_loss_ratio > 2)
- PriceChange prediction: f1_score: 65-85%
- LSTM pure price prediction error: from 0.1% to 4.77% for 1m and 5m candles
3. Maximum/average prediction error (in %):
- max prediction error: 5%,
- median prediction error: 0.2-0.3% for 1m and 5m candles.
4. Ability to predict “x” data points with acceptable accuracy:
Depending on the target, our models generate predictions for the next 4, 8 and 16 data points.
5. Frequency of retraining required to maintain acceptable accuracy:
- Once per 2 weeks for 1h candles.
- Once per week for 15m candles.
6. Trading Strategy Metrics – example for the test of selected stocks portfolio, 10 years:
- Total Trades 68000
- Average Win 0.29%
- Average Loss -0.19%
- Compounding Annual Return 11.65%
- Drawdown 4.7%
- Sharpe Ratio 1.66
- Loss Rate 54%
- Win Rate 46%
- Profit-Loss Ratio 1.6
- Profit Factor 1.39