BTC Strategy — Adversarial Validation ReportRSI(14) > 54 · Price > MA155 · 20-day vol < 100% — daily close, long or cash, nothing else.
September 2026 · All results net of 10bps/order (round-trip cost of 19bps measured with live orders on Binance spot). Period: Jan 2018 – Sep 2026 unless stated. Data: Binance BTCUSDT daily & 1-minute klines (4.77M candles).
Buy & Hold over the same window: 25.9% CAGR, Sharpe 0.69, MaxDD -77%, 7.4x. The strategy compounds 3.4× more capital with less than half the drawdown, by holding BTC only 28% of the time.
01 — Annual results
| Year | Strategy | Buy & Hold |
|---|---|---|
| 2018 | -11% | -66% |
| 2019 | +93% | +94% |
| 2020 | +166% | +302% |
| 2021 | +54% | +60% |
| 2022 | -1% | -64% |
| 2023 | +50% | +156% |
| 2024 | +109% | +121% |
| 2025 | +2% | -6% |
| 2026 | +13% | -7% |
The pattern that drives the whole result: in bull years the strategy captures 50–90% of the upside; in bear years (2018, 2022, 2025) it sits in cash and loses almost nothing while Buy & Hold gives back 60–75%. Compounding does the rest: winning years matter less than not losing them back.
02 — Parameter sweep: 360 configurations
Every combination of RSI period (7/14/21), RSI threshold (45–58), MA window (100–200, SMA and EMA), and volatility cap (0.8/1.0/1.2/none) was backtested — plus RSI hysteresis variants and MACD replacements.
| Config | CAGR 5y | Sharpe 5y | Verdict |
|---|---|---|---|
| RSI14>54 · SMA155 · vol<1.0 (production) | 34.7% | 1.30 | RANK 1 / 360 |
| RSI14>54 · SMA125 · vol<1.0 | 32.8% | 1.22 | runner-up |
| RSI14>54 · EMA100 · vol<1.0 | 32.4% | 1.19 | plateau |
| Best hysteresis variant (in 54 / out 48) | 29.0% | 1.08 | KILL |
| Best MACD variant (rank 57/360) | 20.5% | 0.95 | KILL |
The production config ranked #1 of 360 on both the full period and the last five years — and sits on a broad plateau (neighbouring configs score within 0.1 Sharpe), which is what robustness looks like. A knife-edge winner would be a red flag; this is not one.
03 — Timeframes: the same rules, faster
| Bar size | CAGR gross | CAGR net | Sharpe net | Trades/yr | Verdict |
|---|---|---|---|---|---|
| 5 min | -6.7% | -100% | -30.6 | 4,507 | KILL |
| 15 min | 13.6% | -94.8% | -10.0 | 1,537 | KILL |
| 1 hour | 32.3% | -36.1% | -1.37 | 364 | KILL |
| 4 hours | 49.4% | 27.0% | 0.94 | 81 | KILL (decaying) |
| Daily (production) | 45.8% | 42.5% | 1.29 | 12 | OPTIMAL |
The signal structure works at every speed gross — 4-hour bars even beat daily before costs. But each halving of the bar size roughly triples the trade count while shrinking the edge per trade. At 1-hour bars the strategy pays ~7% of capital per year in fees alone; at 5 minutes, ~90%. Only the daily bar leaves the edge intact — and it is also the only timeframe whose rolling Sharpe is improving (0.87 → 1.48 → 1.56 across 3-year blocks) while 4h decays (1.29 → 0.72).
04 — 66 challenger strategies
| Family | Best variant | CAGR net | Sharpe | MaxDD | Verdict |
|---|---|---|---|---|---|
| Production signal | — | 45.0% | 1.32 | -32% | BASE |
| Signal + dip-buying (RSI7<30 in uptrend) | best of 12 | 47.7% | 1.22 | -32% | KILL — edge lives only in 2018-20 |
| Donchian breakouts (daily & 4h) | 55/20 on 4h | 41.5% | 1.21 | -40% | KILL |
| EMA crossovers (16 combos, daily & 4h) | 30×150 on 4h | 41.7% | 1.05 | -63% | KILL |
| Faster RSI variants (RSI7/RSI10) | RSI7>62 | ~40% | ~1.1 | -35% | KILL |
Not one of 66 challengers beats the production signal on risk-adjusted terms. The single variant with higher CAGR (+2.8pp from dip-buying) loses on Sharpe and draws its entire advantage from the 2018-2020 block.
05 — Long / short
| Variant | CAGR net | Sharpe | MaxDD | Verdict |
|---|---|---|---|---|
| Long-only (production) | 46.8% | 1.15 | -24% | BASE (2019-26 window) |
| Long + short when RSI<46 & price<MA155 | 23.2% | 0.66 | -52% | KILL |
| Long + short whenever signal is off | 24.9% | 0.67 | -64% | KILL |
| Short-only in downtrends | -16.1% | -0.54 | -83% | KILL |
BTC bear markets are punctuated by +30% rallies that liquidate slow-signal shorts before the next leg down — even granting shorts zero funding cost. Cash is a better short than a short.
06 — Execution engineering
| Test | Result |
|---|---|
| Execution lag t+1 → t+2 | -14pp CAGR, Sharpe 1.34 → 1.05 — execute promptly, the signal is fast |
| Exit delayed one extra day | -4.4pp CAGR — sell the day after the signal, without hesitation |
| 3-day signal lock (anti-churn) | ±0 — doesn't help, doesn't hurt; not worth the complexity |
| Real round-trip cost (live orders, Binance spot) | 19bps round-trip measured — 9.5bps/order fee, ~0bps spread |
| Cash yield while flat (72% of the time) | +1-2pp/yr (EUR earn) to +4pp/yr (USDC earn) — the only free improvement found |
07 — Calibrated expectations
Sharpe by 3-year block: 2.71 (2015-17) → 1.81 (2018-20) → 1.22 (2021-23) → 1.34 (2024-26). The edge has decayed from the wild early years and stabilised around 1.2–1.3. Forward expectations should be anchored on the last five years, not the full history: ≈35% CAGR, Sharpe ≈1.3, MaxDD ≈-16%. The full-period 45% includes regimes that likely won't repeat.
Verdict
The production strategy survived every attempt to beat it: 360 parameter variants, 5 timeframes, 66 challenger systems, short overlays, execution tweaks — all tested net of real, live-measured costs. Everything lost. The signal as designed — RSI14 > 54, price > MA155, vol < 100%, daily, long or cash — is the optimum of its entire neighbourhood.
The only genuine improvements are operational, not analytical: execute t+1 without delay, keep costs at ~10bps/order, and earn yield on the 72% of time spent in cash. Do not touch the signal.
Methodology: signals computed on closed daily candles only; positions take effect on the next bar (t+1); 10bps deducted per order on every position change; no leverage, no lookahead (one lookahead bug was caught during this research when a challenger printed Sharpe 8 — the base replica check exposed it). Challenger selection is in-sample by construction: 66 strategies were given every chance to overfit their way to a win, and still lost.