I Backtested a 3-Indicator Crypto Bot on 6 Years of Data — Here's What Actually Beat Buy-and-Hold

Most retail trading strategies fail spectacularly when you stretch them across multiple market regimes. But a methodical combination of RSI oversold signals, EMA trend confirmation, and ATR-driven position sizing — executed through OctoBot’s backtesting engine — just produced something worth paying attention to.

The setup is refreshingly transparent. Using OctoBot 2.1.1 inside an isolated Python 3.12 environment, the workflow pulls BTC/USDT daily candles from a fallback chain of six exchanges (Binance, KuCoin, OKX, Bybit, MEXC, Kraken). The strategy logic is dead simple: it only enters long when RSI dips below a threshold (tested at 25, 30, and 35) AND price sits above the 50/200 EMA cloud. Exits are handled by adaptive stop-losses and take-profits scaled to ATR — 2.0x for stops, with profit targets tested at 3.0x and 5.0x ATR.

Here’s where it gets interesting. Instead of cherry-picking parameters, the author runs an 18-combination grid search across the 2019-2023 in-sample period. Each configuration’s performance is measured not by raw return but by excess return relative to buy-and-hold — a metric that automatically penalizes strategies that merely surf a bull market. The best combo is then validated on completely unseen data from 2023 through mid-2025.

The architecture itself is worth noting for anyone who’s wrestled with Colab’s dependency hell. Rather than fighting preinstalled packages, the workflow builds a clean virtual environment with uv, installs pinned OctoBot and OctoBot-Script versions, and patches the report template for Colab compatibility. The backtesting worker runs asynchronously, right-aligning indicator arrays to handle tulipy’s variable-length outputs, and dumps structured JSON results for downstream analysis with Pandas and Plotly.

I’ve seen plenty of backtests that look gorgeous in-sample and crumble the moment they meet fresh data. The out-of-sample validation here is the real test — and the fact that the author explicitly frames this as an overfitting check rather than a victory lap tells me they understand the game. The interactive charts plotting parameter sensitivity, equity curves, and indicator overlays aren’t just eye candy; they surface whether those RSI thresholds actually matter or if you’re just curve-fitting noise.