Strategy Lab Notes
The stories behind the backtests — where each strategy came from, how it was tested, and what survived. Winners and losers alike.
Honestly Mediocre: the Bitcoin Trend Strategy That Passed the Overfit Test and Failed the Quality Bar
A Donchian breakout on BTC H4 made 7.8% a year across four unseen years with textbook retention numbers. We rejected it on absolute quality, and the same rules on H1 lost 60 times out of 60.
We Finally Tested Our First Strategy the Hard Way. The Edge Held.
The Asian-session breakout that started this project was only ever tested in-sample. So we ran it through the full pipeline: a frozen out-of-sample window, a cost-stress rerun, 10,000 Monte Carlo resamples, and a beat-the-baseline check. It survived all four.
Gold Really Does Rise Overnight. We Couldn't Keep a Cent of It.
The 'London bias' says gold's gains accrue between the PM and AM fixes. We measured it on raw tick data and it's there: +$0.90 per ounce per night. Then we tried to trade it, and the nightly costs came to $1.47.
We Reopened a Closed Case With Better Tools. The Case Stayed Closed.
Month-end rebalancing flow at the London 4pm fix, stacked across four FX pairs with the portfolio tooling we didn't have last time. Best shared-parameter cell over five years: 41 cents a trade.
The Weekend Gap Really Fills. Your Broker Opens Too Late to Catch It.
Published statistics say FX weekend gaps fill 70-86% of the time. Our first backtest produced zero trades, and the reason is the finding: the venue quotes from Sunday evening but only trades from Monday. By then the gap is gone.
Half of a Journal of Finance Effect Is Still Real. It Pays Less Than the Commission.
The FX session cycle is documented in top journals: currencies fall during their own trading hours. We measured it on EURUSD tick data. The European leg still works, at 0.75 pips a trade. Costs are 0.7.
We Tested the Best-Evidenced Idea in Finance at the Scale a Retail Trader Can Reach
Trend following works because of breadth, say Man Group and AQR: many mediocre streams, one good portfolio. We ran the experiment with the six markets a retail account actually has. 58 of 60 configurations lost money.
The Bitcoin Overnight Drift Is Documented at 33% a Year. Our Backtest Lost in All 48 Configurations.
A published BTC seasonality earns 33%/yr on exchange data. On CFD pricing, every configuration we tested lost money before commission. The spread is the whole story, and we measured it from raw ticks.
The Tokyo Fix Anomaly: a Real Edge That Still Didn't Make the Cut
We tested the famous 'gotobi' USDJPY anomaly on 7 years of tick data. It's real, and we still rejected it. Here's the whole story, including the bug we found along the way.
The Backtest Claimed +60% in 11 Months. Ours Found Zero Winning Configurations.
A published gold trading strategy came with a spectacular backtest. We retested the same rules on three years of tick data with real costs and swept the complete parameter grid: all 72 combinations lost money. A case study in why single-period, cost-free backtests can't be trusted.
The Carry Trade We Killed in an Hour, Using Only the Broker's Own Swap Table
Before writing a line of code for a carry strategy, we read IC Markets' real swap rates for eight candidate pairs. The broker's cut shaved the fat interest gaps down to pennies, and the fattest pair of all is one you are not allowed to open.
We Gave a Trend System Five Years of Sample. It Gave Back a 1.17 Profit Factor.
First run of our pre-registered 5-year slow-trend window: EURJPY channel breakouts are mildly real, never validation-grade, and the filter that helped Bitcoin made the carry cross worse.
We Transferred an Academic Momentum Effect to EURUSD. The Market's Most Efficient Pair Said No.
Intraday time-series momentum is documented in equities and RUB-USD. On post-2019 EURUSD, the best of 48 configurations made $371 in three years. A clean transfer-test failure, with costs ruled out.
A Gold System With 28 Years of Published Results Met Our Tester. The Gradient Told the Real Story.
A filtered Donchian breakout on gold claims PF 1.85 over 28 years, costs included. Our grid lost in all 36 configurations, but monotonically less as we slowed it down. Where the edge lives, no gate can follow.
The Triple-Swap Strategy, Finally Backtested With Real Swaps: the Carry Is Real, the Trade Is Not
Broker folklore says you can collect three nights of USDJPY swap in a few hours of exposure every Wednesday. We modeled real swaps and ran it. The carry paid every week. The strategy lost anyway.
The London Fix Reversal: a Real Anomaly That Pays Less Than a Savings Account
We tested the WM/R 4pm fix reversal on 7 years of GBPUSD tick data. It passed 11 of our 12 gates and we still rejected it. The one that failed is the one that matters.
All 146 Configurations Lost Money: RSI(2) Mean Reversion Meets Gold
We transferred one of the most replicated mean-reversion strategies from daily equities to gold H1 and optimized it on three years of tick data with real costs. The optimizer found nothing, and the reason why is a lesson in edge geometry.