How to Read a Strategy Log and Spot Overfitting

Why summary return percentages hide fragile strategies

Two simulated strategies can both finish a four-year BTCUSDT backtest ahead of buy-and-hold by 18 percent while behaving completely differently underneath. Strategy A might capture six multi-month bull regimes calmly and sit in WAIT during prolonged bear trends; Strategy B might churn through 190 rapid switches, lose on 70 percent of them, and owe its entire edge to exiting one afternoon before a flash crash.

Reading the trade execution log turns chart admiration into forensic review. Every row documents timestamp, price, action (BUY or SELL), and the exact regime transition, signal trigger, or risk guard responsible.

Four warning signs of overfitting inside an execution log

When you tune lookback windows, confidence thresholds, blend sliders, and guards repeatedly on the same historical span, you inevitably mould rules around past accidents. Scan the execution log for four concrete red flags before trusting any configuration.

Flag any simulation exhibiting these structural weaknesses.

• Single-trade dependency: removing the single best simulated exit causes the entire multi-year curve to underperform buy-and-hold.
• Ping-pong churn: repeated BUY-then-SELL round trips within two or three bars, which real-world fees and spread slippage would quickly erode.
• Rule shadowing: an aggressive hard stop or SuperTrend flip fires 90 percent of exits before the underlying HMM regime probability ever shifts.
• Cliff-edge sensitivity: nudging the HMM window from 90 bars to 100 bars or confidence by 0.05 flips the net backtest outcome from green to deep red.

Combine forecast history cards with out-of-sample checks

Alongside historical backtest logs in Pro, Market Brief AI Plus keeps the last ten timestamped forecast cards per asset—preserving what regime call, confidence percentage, price, and one-line reading the model actually produced at the moment you ran it.

Follow a disciplined four-step audit to separate durable regime persistence from hindsight tuning.

1. Split your Binance pair’s available history mentally into an earlier calibration period and a later untouched validation period.
2. Choose your model, confidence, and at most one or two signals/guards on the earlier span, then run the exact same settings on the later span without retuning.
3. Inspect the Pro execution log across both spans to confirm average holding duration and rule firing mix remain consistent.
4. Review saved forecast cards offline over subsequent weeks to see how live closed-bar regime confidence evolves in real time without hindsight.

Keep your research honest and offline-ready

Because Market Brief AI stores synced histories, trained models, forecast cards, and execution logs locally in the device’s own database with zero cloud accounts, you can review past runs offline on a flight and delete any asset cleanly whenever you finish an experiment.

A clean execution log proves only that a statistical rule behaved consistently on historical data—never that future crypto markets will repeat those transitions. Treat every audit as an exercise in critical thinking rather than trading advice.