AI can generate a thousand strategies. I tell you which three deserve your money.

Day 1

You bring a mandate

Your target return, risk and drawdown limits, markets you can trade, and anything already in your book. We agree the pass/fail bar before any testing starts.

Days 2–6

I test my own ideas against it

I research and test my own strategy hypotheses for your brief. Rules are written down and locked before results are seen, and every candidate goes through strict anti-overfitting checks.

Day 7

You get the deliverables

A short report with a clear verdict (pass, pass with caveats, or fail) for each strategy, the key numbers set against your mandate, honest caveats, and a recommended next step.

Public pension · crisis-risk offset

Alternative risk premia sleeve

Verdict: pass

The brief. A US state pension keeps 20% of its plan in a crisis-risk-offset portfolio. The risk-premia part must have little stock-market exposure, hold up in crashes like early 2020 and 2022, and beat T-bills + 2% after costs.

One-week approach. A seasonal commodity strategy that earns returns unrelated to stocks, plus a rules-based switch into bonds and gold when stocks weaken, and a small slow-trend slice.

Key results. Stock-market sensitivity −0.02 (target under 0.2). Up in Q1 2020 (+0.2%) and 2022 (+4.2%). Beat cash in 19 of 19 years, about 7.6% a year, worst fall −4.9%. Beat T-bills + 2% by 4.0% a year before Oct 2025 and by 0.7% in the test year.

Honest caveats. Measured above cash, the test-year Sharpe is a modest 0.79 (0.68 with double costs and late trades). After adjusting for the many ideas we have tested, the evidence is moderate, not conclusive. The seasonal part trades rarely.

First-loss seeder · tight drawdown

Low-drawdown levered sleeve

Verdict: pass with caveats

The brief. A seeder that levers a manager's capital (up to 10×) wants small, steady losses: a hard stop at −8.5%, a target worst fall of 3–5% including the COVID crash, a strong Sharpe (we used 1.5) and money made every year.

One-week approach. The same risk-premia sleeve, run at very low risk with day-to-day volatility steering.

Key results. Worst fall −3.1% across the period, including 2008 and early 2020. A random-reshuffle stress test puts a bad-case fall at −6.1%. At 2× leverage the worst fall from a peak is −8.1%, but never more than −4.0% below a year's starting value, so the −8.5% hard stop is never hit. Beat cash in 19 of 19 years.

Honest caveats. Sharpe is 1.56, or 1.23 measured above cash, under the 1.5 target. A bad year could still break a 5% loss limit. The evidence adjusted for the number of ideas tested is moderate.

Multi-manager crypto fund · market neutral

Crypto crowding / funding pod

Verdict: promising, not client-ready

The brief. A multi-manager crypto fund wants market-neutral pods targeting about 20% a year net, a Sharpe above 1.5 (less if truly different from what they hold), and little exposure to the crypto market.

One-week approach. A weekly trade across the most-traded coins: buy the coins where futures traders are least crowded into long bets, sell the most crowded, balanced so a bitcoin move hits both sides equally.

Key results. Test period (Oct 2025 – Nov 2026): 20.3% a year, Sharpe 1.33, worst drop −11%, while bitcoin fell 30%. Almost no link to bitcoin (correlation −0.16).

Honest caveats. Sharpe is below the 1.5 bar. Its worst drop in 2021–22 was −33%. After adjusting for the many ideas tested, we can't rule out luck. Weekly trading is slower than this client prefers. We recommend 3 months of paper trading first.

Crowd-sourced prediction tournament

Crypto ranking signal

Verdict: originality not proven

The brief. A crypto prediction tournament wants a daily ranking of about 300 coins, scored on how well it matches each coin's return over the next ~30 days. It pays most for signals that add something new.

One-week approach. A simple rule-based score with no machine learning: coins that are uncrowded in futures markets and have had calm prices rank high.

Key results. Ranking match with future returns of 0.115 in the test period (typical tournament signals score about 0.01–0.03).

Honest caveats. Almost all the strength comes from the well-known “calm coins beat jumpy coins” pattern (0.139 on its own in the test period), so it may add little that is new. The crowding part alone is weak. We tested about 74 coins, not 300, during a falling market. Next step: submit live, unstaked, for 8–12 rounds.

Trader-seeding programme

Darwinex DarwinIA GOLD qualification

Verdict: pass on the test year, with caveats

The brief. To reach GOLD, a strategy needs 8+ months of live history on the platform and, for example, a 1-year return above 20% with return divided by worst fall above 2.5. The platform runs it at a fixed risk level.

One-week approach. The risk-premia sleeve scaled up to the platform's fixed risk level.

Key results. Test year (Oct 2025 – Oct 2026): +24.6% with a worst fall of −5.9% (ratio 4.2). With double costs and late trades: +23.0% (ratio 3.5).

Honest caveats. Leverage: this needs up to about 3× leverage. Capped at 2×, the test-year return falls to +16.3%, below the 20% bar. Consistency: it would have qualified in only 6 of 10 past years, so any given year is likely but not assured. It also needs futures access on the platform and 8+ months of live history there.

Locked holdout period

The most recent stretch of data is set aside and never used to pick or tune anything. It is used once, as a final exam.

Luck test

If you try enough ideas, some look great by chance. The results are adjusted for how many ideas were tried overall.

Reshuffle test

The returns are randomly reshuffled many times to check the result is not a fluke of one lucky sequence, and how bad a bad run could get.

Delay and cost stress

Every trade is filled a bar late and trading costs are doubled. A strategy that only works with perfect timing or free trading is dropped.