Faber, Halloween, Turn of the Month, Turtle: an AI Tested 1,306 Variants and None Survived

The most quoted trading rules come from academic papers or market legends: Faber's 10-month timing, the Halloween effect ("Sell in May"), the turn of the month, 12-month momentum, the Turtle 20/10 breakout. Each has a brilliant track record… before it was published. The question that matters is what they do afterwards.

On October 12, 2026, we handed that question to an AI. DeepSeek (deepseek-flash, through opencode), connected over MCP to the Strategy Arena Lab, had one instruction: find a strategy that survives the judge. It made 97 tool calls, explored 7 families, and the Lab counted 1,306 trials.

Result: no survivor.

Research on past data, not investment advice.

The rules of the game

The AI did not pick its own referee. The Lab's judge is the same for everyone:

• Every trial is counted. A variant tested inside a sweep counts, even if the AI never shows it. The deflated Sharpe is computed with that real number.
• One sealed period per market. For published rules, the judge reserves a period after their publication, never seen during tuning.
• The baseline is buy-and-hold on the same scope, same fees.
• "Survives" requires a gain above buy-and-hold, every check passed, and at least 30 independent events in the test period.

The 7 families

Family by family:

• Faber (10 months). Before publication (1926-2007), the rule had a 42% maximum drawdown against 71% for the market. In 2012-2026: +316% against +610%. Above all, a monthly rule produces too few events: 11 trades, a single independent event. It can never reach the 30 required. It is not wrong; it is unjudgeable at that scale.
• Halloween effect. Before 2013: +497% against +486%, with a 33% maximum drawdown against 62%. In 2015-2026: +106% against +152%, with the same maximum drawdown (35.7% on both sides). SPA p = 1.00: the gap is compatible with chance.
• Turn of the month. The only rule that clears the 30-event bar (156 out of 156 trades). It still fails: −12.3% against +176.1%, and 68% of Monte Carlo draws end in a loss. Before 2008 it was already a low-return profile: +93% against +529%.
• 12-month momentum on EUR/USD. Even before publication (1999-2012), no edge: +13.7% against +13.0%. In 2018-2026: −10.4% against −3.9%, 91.9% of Monte Carlo draws in a loss.
• Turtle 20/10 on EUR/USD. The textbook case of overfitting: +31.9% in training, then −42.1% in 2013-2026, with a 50% maximum drawdown against 31% for buy-and-hold. 99.9% of Monte Carlo draws in a loss. The 2N stop fires constantly in a trendless EUR/USD.
• Volatility compression on BTC 4h. The most explored family: 1,300 variants swept. Over 8 years, the best setting made +2,606% against +2,450% for buy-and-hold, with a 43% maximum drawdown against 84%. At the judge: 1 walk-forward fold out of 4 above buy-and-hold, deflated Sharpe of 3.9% with 1,301 declared trials, SPA p = 0.44. Mirage.

The twist: they protect, but they miss the rallies

It would be wrong to say these rules "do nothing". They do something measurable: they reduce maximum drawdown.

And in bear markets they beat buy-and-hold: +40.6 points for the turn of the month, +24.8 points for EUR/USD momentum. In its test period, the Halloween rule loses 50.5% across the bearish stretches, against 71.2% for the market.

The price is paid the other way. In the bullish stretches of that same test period, Halloween makes +152.9% while the market makes +388.6%. A rule that is out of the market half the time misses half the rallies. And in 2012-2026, buy-and-hold was very strong.

It is the same finding as our test of 25 strategies across 9 markets, where none beat buy-and-hold in a bull market: a real defense is not an edge as long as what it costs in rallies exceeds what it saves in declines. If you want to sleep through a crash, these rules have value. When you compare them with buy-and-hold, they lose.

The flaw the AI found in our Lab

The session also exposed a weakness of the Lab itself. For volatility compression, DeepSeek swept 2023-2026 without knowing that this window was the judge's sealed period. When it asked for the verdict, 100% of the test candles had already been seen by its own exploration. The judge detected it and confirmed nothing, but the warning came too late.

The fix is in the next Lab version, and it was checked the following day on the installed Lab: a backtest that overlaps a sealed period gets the band_already_opened warning, with the safe range to explore, before anything is spent. No refusal: the AI stays free; it simply knows what it is burning.

Another limit reported by the AI: there is almost no fresh crypto period left to judge. Every crypto test window in the Lab was already opened, which is why the weekend filter is "not judged".

What this result says, and what it does not

• It says that, judged on their post-publication period, fees included and trials counted, these five famous rules do not beat buy-and-hold on the markets tested.
• It does not say they will never work elsewhere. Every new market, every new setting would be a new counted trial.
• 1,306 trials are not 1,306 ideas: 1,300 come from a single family. The five published rules were each judged once, untuned, as published.

On connecting an AI to this judge: DeepSeek, Claude, Codex: plugging an AI into the backtester. On why a searching AI always finds "something": 350 autoresearch sessions on noise.

Run your own hunt

• The free Windows Lab: your AI, your GPU, the same judge. The MCP guide explains how to connect Claude Code, Codex or opencode.
• The free Quick Check: paste your Pine strategy and see whether it holds up against buy-and-hold, fees included.

Key takeaways

• 7 families, 1,306 counted trials, 97 tool calls: no strategy survives.
• Faber and Halloween: inconclusive (below the market, or too few events). Turn of the month, 12-month momentum, Turtle: mirages.
• Turtle made +31.9% in training and lost 42.1% in the test period.
• These rules cut maximum drawdown by 1.3 to 3 times and win in bear markets, but they miss the rallies.
• The AI exposed a Lab weakness (exploring inside the sealed period); the warning now exists.

Research on past data, not investment advice. No past result guarantees a future result.