Once a week LiTA trains a random-forest classifier (300 trees, max depth 6, class-balanced) on weekly on-chain, macro and technical features and predicts the direction of the next ~14 days: BUY or SELL, with the winning class probability reported as confidence. A 6-week minimum hold stops the signal from flipping every week, and a circuit breaker forces SELL when price has been 40%+ below its all-time high for 3 consecutive weeks. Each live signal is scored 14 days after issuance against the actual closing price. LiTA is in a paper-trading validation period; no capital is traded.
1Backtest figures come from the Phase 6F walk-forward equity backtest (104-week train / 13-week test / 13-week step), resampled weekly and re-based so every strategy starts on the same date. They are re-computed only when btc_equity_backtest.py / eth_equity_backtest.py run; the date range is shown above the chart.
2“After fee” applies the per-trade fee configured in config/fees.py when the backtest ran (Coinbase worst-case policy). the fee is charged on every position change. After-tax applies short- or long-term capital-gains rates on realised gains. SELL weeks are modelled as short positions without funding costs; weeks where the model’s confidence is below its threshold are held flat (0% return).
3Live hit rate counts a signal correct when its direction matched the sign of the actual 14-day return. Small samples are not statistically conclusive.
4LiTA is in a paper-trading validation period. Nothing on this page is financial advice.