ML models showing high win‑rates
AI agents processing 1,200+ input points are being touted to generate signals with ~78% win‑rates across assets—claims to verify but noteworthy for systematic‑strategy builders . On prediction markets, a Conv1D+LSTM model applied to 38 variables (RSI, MACD, etc.) reportedly hit 59–78% win rates with 100+ trades/month—specifics worth backtesting before allocating capital .
Validating these social-media claims demands a chronological in‑sample → walk‑forward → out‑of‑sample protocol; the AlgoXpert IS–WFA–OOS framework formalizes that three‑stage approach. (arxiv.org) Adjusting for multiple testing and selection bias requires data‑snooping corrections such as White’s Reality Check and bootstrap-based tests, methods used in major evaluations of technical rules. (kevinsheppard.com) High‑frequency signal sets are especially sensitive to execution costs because spreads, slippage and market impact frequently convert simulated profits into losses; backtesting guides emphasize modeling commissions and dynamic slippage. (backtestme.com) Win‑rate alone omits payout size and risk profile, so professional evaluations prefer expectancy, profit factor, Sharpe ratio and maximum drawdown when judging a strategy’s economic value. (atlantictrading.com) Architectures fed with thousands of inputs confront the curse of dimensionality, which increases spurious correlations; the literature recommends PCA/autoencoders, feature selection, or regularization to reduce overfitting. (towardsdatascience.com) Assessing whether an observed win proportion is real requires adequate sample size and binomial confidence intervals; practitioners commonly treat on‑the‑order‑of‑hundreds of trades as a more reliable floor and compute Clopper–Pearson or normal‑approximation CIs. (gainium.io) Reproducibility hinges on publishing code, data slices, and deterministic backtest pipelines because ML finance research has documented leakage and replication failures when those artifacts are omitted. (jmlr.org)
