The bot returned 4.33%. Doing nothing returned 127.77%. The useful part is why.

A developer writing on dev.to published a three-year out-of-sample result for a crypto trading system: +4.33%, against +127.77% for buying Bitcoin once and holding it ^[1]. That is a gap of 123.44 percentage points ^[2], and the post is worth reading because most of its length goes to the four ways the same strategy looked profitable before the test harness was rebuilt ^[3].

The thing under test is not exotic. About 4,000 lines of Python: a strategy engine, a backtester, a walk-forward validation harness, a paper-trading bot with Telegram alerts, a dashboard, and 130-plus tests in CI ^[4]. The rule is a Donchian channel breakout, buy 20-day highs and exit on a trailing stop, run on exchange data through ccxt ^[5].

Lie one was look-ahead. The old code computed a signal from a candle and then bought at that same candle's close ^[6]. The fix was structural rather than a patch: one engine enforces the ordering for every strategy, and a test feeds in a strategy that deliberately peeks one bar ahead and asserts it cannot profit ^[7]. That is the part any team can copy tomorrow. The bug becomes a permanent adversarial test, so the next contributor cannot reintroduce it quietly.

Lie two was fees, which the original backtester set to zero ^[8]. Kraken's taker fee is 0.26% per side ^[9], or 0.52% for a round trip ^[10]. Re-run with honest costs, the same signal produced: on daily bars, 14 trades, 1.3% of account per year in fees, +4.33%; on 4-hour bars, 61 trades, 11.4% in fees, -6.01%; on 1-minute bars, 52 trades in 45 days, 23.5% in fees over those 45 days, -25.18% and not a single winning trade ^[11]. The signal was identical at every speed ^[12]. On 1-minute bars the author puts the fee at roughly six times the average move being targeted ^[13]. Even the surviving configuration is thin: 1.3% a year in fees against roughly 1.44% a year net implies fees ate about 47% of gross return ^[14]. One thing the post leaves unexplained is that the 4-hour run has about 4.4 times the trades of the daily run but about 8.8 times the annual fee drag ^[15], which anyone reproducing this should reconcile against window length and position sizing.

Lie three was tuning and testing on the same data. The +4.33% comes from walk-forward analysis: optimise on a training window, trade the next window the optimiser never saw, roll forward, and score only unseen days across five years of data ^[16].

Lie four was luck. A permutation test shuffled daily returns 200 times and re-ran the strategy on each scrambled history; the strategy beat 97.5% of the shuffles, p = 0.025 ^[17]. The author's own reading: the edge is real, small, and smaller than the fees ^[18].

Three priors died on the same bench. Seven summed indicators finished last against a one-rule strategy, -2.4% versus +3.0% ^[19]. Equal sleeves of BTC, ETH and SOL produced a portfolio Sharpe of 0.30 against 0.38 for BTC alone ^[20]; the author kept the feature anyway, on the grounds that a bench that spares its author's ideas is worthless ^[21]. The one defensible win: through a stretch where Bitcoin drew down 53%, the bot's worst drawdown was 4% ^[22].

Watch whether the anti-leak test stays green as strategies accumulate, and whether the drawdown behaviour survives a live down leg rather than a scored one.