The honest backtester that rejected two strategies, net of cost

Crypto strategy projects usually ship the exciting part first, a strategy that backtests beautifully, then loses money live because the backtest never survived look-ahead, real fees or an out-of-sample test. quant-ring proves the discipline before trusting any strategy: one NautilusTrader engine runs the identical code in backtest and live, so there is no separate live version to drift out of sync with what got tested, and evaluate.py judges only on data a strategy has never seen.

evaluate.py enforces an out-of-sample split, walk-forward validation across six windows, net-of-cost Sharpe, drawdown and an overfit detector, all covered by 24 tests. Two fighters went through it on two years of Binance BTCUSDT hourly data spanning a full cycle, a trend signal on an MA cross and a mean-reversion signal on a z-score, each judged at two cost levels, a 0.3% round trip and a stressed 0.4%. The public face is the Honest Backtester, a web tool where the browser runs its own calculation and matches the Python backtest to the digit.

Both fighters failed all four runs: the trend strategy lost 67.8% to 75.5% net of cost, mean reversion lost 52.4% to 66.7% and the published table lists every row rejected. No strategy has passed the out-of-sample bar yet, so nothing here trades real money. A negative result counts as money saved rather than money lost live. This is an internal research project, not investment advice. The next step is one new hypothesis under the same discipline, not a pile of parameter variations on the same data.