Test protocol v1

Window

One window 2025-11-01 → 2026-11-01 (end excluded) — the latest full year. It almost fully contains a bear market: BTC fell from its 126,196 high (2025-10-06) to 57,802 (2026-07-01), −54%. Data: Bybit candles.

Pair list

A fixed list: top 80 by average daily turnover over 30 days before 2025-11-01, coin age at least 90 days, exclusions — only our live bot's blacklist. Separate lists for spot and futures. The list is chosen on the first day of the window and never changes during the year, so every strategy trades the same coins and nobody “peeks” at which coins will rise later.

Spot: 80 pairs

SHA-256 bbadaad9198247af4fb7843c811916452f0ce5f43b8277924199193517499234 · download JSON · source selection file, SHA 71dc6e2fa139

Futures: 80 pairs

SHA-256 ca39dbebf3518b9e5a58e76829f48ad0023e22b5c9cba5d5ab676a0f6f33ca17 · download JSON · source selection file, SHA d3f78c30ac3c

Run settings

Every run on the site shows the SHA-256 of the strategy code, run config, data (the list of candle files with their SHA) and pair list — enough to repeat the run.

Metrics

• Profit — sum of closed-trade profit as a percent of the starting wallet (fixed stake, no reinvesting).
• MTM drawdown — every day open positions are revalued at the close of the day's last 5-minute candle (UTC), shorts with the opposite sign; the largest fall of equity from its peak is taken. Stricter than Freqtrade's closed-trade drawdown.
• Profit/drawdown — the main rating metric; “n/a” when drawdown is zero or there are no trades.
• Win rate and profit factor (gross profit / gross loss).
• Monthly returns — growth of MTM equity over the month, as a percent of the starting wallet.
• Exit reasons are grouped by type (profit-target pullback, indicator signal, stop, bad trade, window end) and entry mode; the full reason of every trade is in the CSV.
• “When it made money”: day regime — BTC 7-day change above +3% is up, below −3% is down, otherwise flat; a long trade is one held over 7 days.

Lookahead checks

Freqtrade's standard check (lookahead-analysis): for a sample of trades the backtest is repeated on one pair with data cut right after the entry or exit. If the trade does not repeat in the cut run, the decision depended on future candles. We check January 2026 and February–April 2026.

Strategies that keep state between runs get false flags from this check. So we re-run every flagged trade ourselves, each run in a clean folder: A — all pairs, B — one pair, C — one pair up to the exit, D — one pair up to the entry. The verdict for each trade is on the strategy's Results tab.

• not found — Freqtrade flagged nothing.
• caveat — Freqtrade flagged trades but our re-check did not confirm lookahead; the strategy keeps its place in the rating.
• found — lookahead confirmed in at least one trade; the strategy gets a red mark and moves to the bottom of the rating.

The second check (recursive-analysis) looks at whether indicator values change with the warm-up length. If they do, we check whether the trades themselves change with a different warm-up.

Verification levels

1. Protocol backtest — any strategy with a verified license.
2. Plus “Explanation”: entry schemes, conditions, protections and exits analysed by Glassbox (for now the NostalgiaForInfinity family).
3. Plus forward testing on a paper or live bot (percentages only) — the project's next stage.