Blog

Long-form notes from the research lab.

Methodology pieces, fee-drag findings, calibration retrospectives, and whatever else the data shows. Posts include the negative results, by design. RSS.

• May 19, 2026·4 min read·bandits, thompson-sampling, methodology, walk-forward, genetic-algorithm

  Thompson Sampling over a gene-regime grid: why I picked it over UCB

  I needed an explore/exploit policy for a research lab that tests genetic-algorithm offspring against historical regimes. Here is why Thompson Sampling beat epsilon-greedy and UCB for that specific shape of problem, and the three failure modes I am watching for.

• May 4, 2026·6 min read·backtesting, deflated sharpe, methodology, honesty

  I audited my own strategy approval logic. 549 of 595 'approved' strategies were noise.

  My research engine had approved 595 strategies for forward testing. When I checked the actual gate logic against the design doc, 99% of those approvals had Deflated Sharpe probabilities flagging them as flukes, and 40% had negative walk-forward P&L. The gate was reading from the wrong field. This is what I found and how I fixed it.

• May 1, 2026·6 min read·walk-forward, parameter-search, take-profit, postmortem, methodology

  Same entry, four take-profits: how a 4% target turned a profitable signal into 0/8

  We had four sibling strategies with identical entries and only the take-profit changed. Three were profitable in forward test. The fourth went 0 for 8 with every exit on the timeout. Here's what the data showed and why our parameter search should have refused to generate that variant.

• Apr 30, 2026·13 min read·backtesting, fee drag, paper trading, methodology

  I paper-traded 22 popular crypto strategies on real fees for 10 days. 16 of them lost money. Here's the data.

  Across 26,765 paper trades, the average per-trade P&L is -0.078% and cumulative is -2,081%. The 6 strategies that survive are all in one indicator family. The 16 that don't share another pattern. Real-fee data, not zero-fee marketing numbers.