Most sites show you winners. We show you the failures first.

The claim

The Watchdog ran on a $6,000 real-money BTC allocation from 2026-04-12 to 2026-04-22, when it was paused due to a chart-structure concern at the BUY trigger ($78,880 at ascending-channel resistance). Six days later, the v2.1 multi-benchmark formal recompute showed Sharpe 0.35 + DD-ratio 1.38 vs S&P 500 — all three v2.1 paths fail strict (Path 1 return-superior unverified, Path 2 risk-adjusted FAIL, Path 3 diversification sleeve FAIL with Sharpe below the 0.8 threshold). Same evening: adopted the all-paper policy. No bot graduates to real money until ≥6 months of forward-validated proof. Watchdog demoted Tier 1 → Tier 2. Tier 1 now contains only Basis Sentinel (Path 3) and Alpha Hunter (Path 1).

Root cause

Two compounding factors. (1) Watchdog's 'Tier 1' classification was a v1-era artifact: it cleared the per-trade W/L test (3.44:1) but failed Walk-Forward (25%). It ran on real money under user discretion — a historical decision, not a graduation under the validation framework. (2) The v2.1 multi-benchmark switch from BTC HODL → S&P 500 made the borderline obvious: against BTC HODL, Watchdog's drawdown protection was real (-34.7% vs HODL -77%); against S&P 500, Sharpe 0.35 and DD-ratio 1.38 mean it underperforms a passive index fund risk-adjusted. Calling it 'real-money-eligible' under v2.1 strict would have been special-pleading.

Lesson

Backtest validation tells you a strategy COULD work; it doesn't tell you it DOES work in the actual conditions of the next six months. BBR is two weeks old — there is no forward track record for any bot. The honest move is to require ≥6 months of live paper-tracking with maintained Sharpe + bounded MaxDD + forward beat-rate vs S&P ≥50% before any real-money allocation. Real-Money Graduation Criteria are now codified on /methodology — first eligibility window 2026-10-28. The framework now stands on its own: backtests choose candidates, forward results promote them.

The claim

The Contrarian inverted textbook RSI on BTC: LONG when RSI(14) crossed above 70 (textbook says SELL), CASH when RSI crossed below 30. The bot card claimed +711% over 8 years vs HODL +473%. Validation suite revealed two structural problems: (1) the deployed config (P=14, 70/30) was the WORST point in BOTH the period sweep AND the threshold sweep — adjacent (P=7, 75/25) would have returned +30,000% over 11.6y vs the deployed +2,122%. (2) When tested under WF-on-params with EVERY plausible config including the in-sample winner: NO config beat HODL OOS more than 38% of the time (target ≥60%). Per-trade W/L was 4.92:1 (HIGHEST in entire BBR suite) but couldn't compensate for the directional underperformance.

Root cause

The contrarian thesis is REAL — at every threshold pair tested, inverted RSI beats textbook RSI by hundreds of percentage points. But on BTC over 11.6 years, no specific calibration of the inverted strategy reliably beats HODL across rolling windows. Best-case: (P=5, 75/25) at 38% beat-rate, avg-excess −50pp. The strategy CONCEPT works (textbook RSI loses on BTC), but as a standalone bot it can't be calibrated to consistently outperform buy-and-hold. The +711% over 8 years was a window-cherry-picked artifact (replicated cleanly: +637% on 2018-2026); on the full available history (2014-2026), the same deployed config loses to HODL by −14,757pp.

Lesson

Methodological lesson added to BOT_VALIDATION_STANDARDS: when a bot fails Test 1 (Walk-Forward), don't just report the failure — run a SECOND Walk-Forward over the parameter sweep itself. This separates 'calibration is wrong' (a fixable problem) from 'concept doesn't generalize' (a structural problem). The Contrarian was the latter. Re-calibrating to the in-sample winner would have been pure hindsight bias — the in-sample best (P=7, 75/25) only marginally outperformed (35% vs 30% beat-rate, both FAIL).

The claim

Hypothesis: BTC outperforms when US Dollar Index (DXY) is weak AND S&P 500 (SPX) is strong. Tested as standalone signal and as filter overlay on existing momentum bots. Standalone: only LB=90 passed 2/3 walk-forward, isolated peak (LB=20/30/60 all fail). As overlay on Tactician M-30: dropped Tactician's return from +2,010% to +97-575% — strictly worse on every macro lookback tested.

Root cause

DXY+SPX filter kills too much BTC exposure during bull-market phases. Many of BTC's biggest up-moves happen in 'wrong' macro regimes (e.g., BTC rallied in 2020 while DXY also strengthened initially). The macro signal is too coarse and slow vs BTC's own cycle dynamics. Single-condition variants (DXY weak only, SPX strong only) also fail — the combined filter is the problem.

Lesson

Equity-style macro overlays (DXY/SPX) don't generalize cleanly to crypto. BTC has its own cycle dynamics that dominate macro signals on multi-year horizons. Macro overlays may still work for risk management (reduce position size, not gate trades), but as binary trade-or-not filters they destroy more alpha than they save.

The claim

Hypothesis: Trend-following on Gold Miners ETF (GDX) generates alpha vs GDX HODL. Tested 4 lookbacks (M-20, M-30, M-60, M-90). All variants underperformed GDX HODL on full period. Walk-forward 0-1/3 windows.

Root cause

Gold and gold miners don't have the kind of persistent retail-driven momentum cycles that crypto has. Gold price is dominated by central-bank purchases, currency hedging flows, and macro-inflation expectations — slower-moving institutional drivers. Trend-following requires predictable persistence; gold doesn't deliver it on the daily-momentum timeframe.

Lesson

Strategy classes that work brilliantly in one asset (BTC momentum) don't necessarily port to other assets (GDX momentum) even if the assets have similar volatility profiles. The market structure matters more than surface-level volatility.

The claim

The Scout (Hash Rate Momentum) was downgraded from PAPER to BACKTEST tier on 2026-04-15 when full-period analysis showed it loses to BTC HODL by ~1,572 percentage points over 6 years. We kept it visible on /bots with a BACKTEST badge as an 'educational warning example.' On reflection: the educational value is fully captured by its individual post-mortem article. Keeping the bot on the live grid implied 'active project' which it isn't.

Root cause

Multi-metric review: Scout fails on every single metric (negative Calmar, negative Sharpe, beats HODL in <50% of rolling windows, 1/3 walk-forward). The 'warning example' positioning was defensible but the ongoing Hero-grid presence created cognitive overhead for readers without adding insight beyond what the existing post-mortem article already delivers.

Lesson

Failed strategies belong in the post-mortem ledger, not in the active lineup with caveats. A clean Hero grid showing only currently-tracked bots reads more honestly than a mixed grid with status badges trying to explain why one entry is dead.

The claim

MVRV (Market Value / Realized Value) is a well-known crypto cycle indicator. Tested 6 threshold combinations: SELL when MVRV > {2.5, 3.0, 3.5, 4.0}, BUY when MVRV < {1.0, 1.2, 1.5}. Returns ranged from +177% to +1,162% over 6 years vs HODL +901%. Maximum drawdowns -35% to -67%.

Root cause

MVRV thresholds trigger so rarely (~0.01 trades/week, basically once every 18-24 months) that by the time the signal fires, the cycle has often already turned. Walk-forward 1/3 windows max — no robust regime detection. The single 'good' configuration (3.5/1.2 → +1,162%) is isolated; neighbors don't agree.

Lesson

MVRV may be useful as a regime overlay for OTHER strategies (e.g., reduce position size when MVRV > 3, regardless of momentum signal), but as a standalone strategy it's too coarse. The data is good; the threshold-based execution is wrong.

The claim

Tested z-score mean-reversion of the BTC/ETH price ratio across 9 parameter combinations (threshold 1-2σ, lookback 30-90 days). Logic: when BTC/ETH ratio is z-score extreme positive (BTC overpriced vs ETH), buy ETH; reverse for opposite. Classic equity-pairs setup applied to crypto.

Root cause

ALL 9 combinations failed walk-forward (0-1 of 3 windows). Drawdowns -88% to -92%. Returns -33% to +572% — most below HODL by huge margins. The BTC/ETH ratio doesn't mean-revert in crypto's regime-shifting markets. ETH outperformance from DeFi summer doesn't snap back; BTC outperformance during bear flight-to-quality doesn't snap back. Pairs trading requires stable fundamental relationships; BTC and ETH have shifting narratives that defy mean-reversion.

Lesson

This is the OPPOSITE of what the Rotator does, and the Rotator works because momentum persists. Trying to fade the same signal creates an anti-edge. Strategy classes that work in equities (mean-reversion of correlated pairs) don't necessarily port to crypto.

The claim

Attempted to revive the failed Hash Rate Momentum signal (original Scout's strategy) by adding a regime filter: only trade when BTC > N-day SMA. Tested 9 combinations (HR lookback × SMA lookback). One combo passed walk-forward 2/3, but it was an isolated peak with neighbors at 0/3. Even that 'winner' returned +388% vs HODL +544% — losing to HODL.

Root cause

Same Sharpshooter signature: only HR-60d × SMA-150 passed 2/3 WF, while neighbors (30d/150, 90d/150) scored 0/3. Single isolated configuration, not a robust plateau. And it still loses to HODL on the full period. Hash Rate appears to be a confounded signal — moves in concert with BTC price (miners scale to match revenue), so it offers minimal independent information for timing.

Lesson

A failed strategy with a regime overlay can occasionally LOOK like it works in 1 of 3 windows, but the 4-test stack (continuous + WF + robustness + benchmark) catches it. This confirms: Hash Rate is not an actionable trading signal, even with regime gating.

The claim

On 8 years of BTC daily data, Momentum-14d returned +2,053% (vs HODL +654%), beat HODL in all 3 walk-forward windows, and traded at 1.0 trades/week — looked like a clean upgrade to The Tactician (Momentum-30d). Almost deployed.

Root cause

Parameter sweep showed M-14 was an isolated spike: M-13 returned +437% with 1/3 WF, M-15 returned +768% with 1/3 WF. The 14-day result was 3-5x better than its immediate neighbors, with no plateau supporting it. Classic signature of noise-overfitting — one specific parameter value happened to align with random patterns in the historical data, no underlying market structure that 14 days uniquely captures.

Lesson

Walk-forward and parameter robustness are complementary tests. A strategy can pass 3/3 walk-forward by accidentally aligning with random patterns in each sub-period. The sweep test catches this. After running 17 lookback variants, finding one that beats every other test is statistically expected — not proof of edge.

The claim

We tested two combinations of the Watchdog (cycle filter) and Tactician (30-day momentum) bots. Option B: 50/50 independent capital split. Option C: Watchdog-gated (Tactician trades only when Watchdog=LONG). Continuous-run results looked excellent — Option C returned +218% vs Watchdog-solo's +202% with lower drawdown. Looked like a strict upgrade.

Root cause

Split the data into 3 non-overlapping windows. Neither combo reliably beats Watchdog-solo. Option C: wins 1/3 windows on return (loses W2 by 7pp, W3 by 17pp). Option B: wins 1/3 windows on return AND 1/3 on drawdown. The +218% edge was W1-dominated — a bear-market artifact that doesn't generalize to bull regimes.

Lesson

Continuous backtests can look great and still fail walk-forward. The 3-window check is the truth serum. Both combos provide marginal drawdown protection at meaningful return cost in most regimes. Neither is a clean upgrade over Watchdog-solo. Signal correlation (~70%) means the combos don't diversify enough to overcome the fees and tax friction they add.

The claim

Scout (Hash Rate Momentum) was originally promoted as a live paper-trading bot. Honest backtest over 6 years: +575% bot vs +2,147% HODL. Scout loses to buy-and-hold by a ~1,572 percentage-point margin over the full window.

Root cause

Scout performs well in specific regime slices (2022-2024) but terribly in others (2020-2022). Averaging hides the fact that it systematically underperforms during strong trending markets, which make up most of the BTC 6-year record.

Lesson

A strategy can look great in a cherry-picked window and still fail the honest benchmark. BACKTEST-only label keeps the bot visible on the /bots page as an educational artifact, not a recommendation.

The claim

An attempt to rebuild the Scout around a different signal: Donchian 20-day breakout with 50-day SMA regime filter, 15% TP / 5% SL, 10-day time stop, positive-funding requirement. Continuous backtest 2020-2026: +156% over 71 trades. Looked clean.

Root cause

Walk-forward split the 6 years into 3 equal windows. Scout v2 lost to HODL by 346pp in W1 (biggest BTC bull in history), lost by 32pp in W2 (mid-cycle recovery), won by 10pp only in W3 (quiet consolidation where HODL itself was weak). Fixed TP/SL ratios too tight for BTC volatility — the bot missed the asymmetric upside that makes HODL hard to beat.

Lesson

A continuous backtest can look presentable and still fail walk-forward. This is the third strategy in a week (after Options B and C of the Watchdog+Tactician combo) where the 3-window check killed a result that seemed promising on the full period. The Scout family — mid-frequency trend-following on BTC daily — appears structurally disadvantaged vs HODL in asymmetric markets.

The claim

Inverse-vol-weighted Momentum Top 20 US stocks. Backtest +412% vs SPY +212%, -41% MaxDD (vs Alpha Hunter's -53%). Ran one paper rebalance with 20 positions. Looked promising.

Root cause

Same asset class (US large-cap stocks) and same signal family (momentum) as the existing Alpha Hunter. It would have diluted the bot lineup with a near-duplicate strategy that readers couldn't distinguish at a glance. The 'same alpha, less pain' pitch was real but marginal, and it violated the unique-angle filter we apply to new bots.

Lesson

Not every backtest-positive strategy deserves a production slot. Quality bar: unique asset class OR unique signal family. Alpha Hunter Light failed both tests.

The claim

Classic RSI(14) oversold-overbought bot. Tested on 6 years of BTC daily data. 58 trades. 28% win rate. Lost money in every parameter configuration tested.

Root cause

RSI is a mean-reversion indicator. Bitcoin does not mean-revert on daily scale. It trends. Trend-persistence dominates mean reversion in the base rate of BTC price action. RSI buy-signals (oversold bounces) are wiped out by continuation.

Lesson

Popular doesn't equal working. An indicator can be ubiquitous in financial media and still fail a proper test. RSI survives because it looks right on cherry-picked charts. On the full history, it's negative-expectation.