All articles
PAPER · PRE-REGISTERED
Research note
A published anomaly, sample ending 2019: intraday losers beat intraday winners in the last half hour. We tested it on 2020–2026 under a pre-registration written before the run, on S&P 500 stocks minus the heavily traded names we had already seen. It is still there. It depends on the window, the end point and the sample, and the paper shows where.
MACRO · 2008
Research note
On the day the S&P 500 set its 2007 record, the six big homebuilders were already down 62% and $283 billion of short-term funding had vanished. I lined up 38 prices by the date each one peaked. The crash rolled outward from mortgages for 27 months, and the order is measurable.
PAPER · SSRN
Research note
A text model can read headlines beside a trading system the way a thermometer sits beside a patient: it reads, it never prescribes. Six gates it had to pass first, the one line it never crosses, and all 22 times it was used, failures included.
PAPER · SSRN
Research note
Six implementation defects caught by replaying live days through the backtest: a thin data feed, adjusted prices, a runaway sizing step, a UTC clock, missing inputs, dead config. None showed in the results. Each now has a test.
MACRO · JAPAN
Research note
In 2024 a crowded yen trade turned a small Tokyo hike and one weak US jobs report into the Nikkei's worst day since 1987. It unwound in two weeks, then got rebuilt. How the trade works, why the exit is violent, and what the September 2026 Fed and BoJ meetings did to it.
PAPER · SSRN
Research note
The validation machinery behind every note on this site, written up properly: a failure-mode taxonomy, a Monte Carlo of both error types, and a new pre-registrable criterion. Free on SSRN, simulation code included.
THESIS · COMPLIANCE AS CODE
Research note
A compliance sign-off tells you the system passed on the day it was checked. Nobody measures what it does the rest of the year.
REJECTED · STOP-LOSSES
Research note
The worst months and the best months are usually the same month. Every threshold we tested took the loss and then sat out the recovery.
INFRASTRUCTURE · DATA STACK
Research note
A one-person lab doesn't need twenty tools. Here are the five we kept, and why the license file mattered as much as the benchmarks.
INFRASTRUCTURE · AGENT DATABASES
Research note
It held up fine with humans. Agents write faster, retry more, and don't wait their turn. One database per agent fixed the corruption, and made backtests reproducible as a side effect.
INFRASTRUCTURE · OBSIDIAN + AGENTS
Research note
It read my vault and made a well-sourced case for a thesis I'd killed two months earlier. I nearly acted on it. The fix was a handful of documents — 2% of the words — that tell any reader how to read the rest.
SHELVED · REAL EFFECTS
Research note
The effect was real for years and there was a reason for it. We still shelved it. What decided was capacity, turnover, and costs.
REJECTED · SHORT SIDE
Research note
If the top of a ranking works, the bottom should fail by the same amount. In our backtests the two legs turned out to be different signals, and only one of them worked.
REJECTED · DIVERSIFICATION OVERLAYS
Research note
Five overlays, added one at a time. Each made the backtest worse, and none of them brought information the base signal didn't already have.
REJECTED · EVENT-NIGHT FILTERS
Research note
Earnings, announcements, big news — we filtered them out. Most of the performance went with them.
REJECTED · LONGER HOLDS
Research note
Longer holding periods improved every backtest we tried them on. On held-out data the improvement disappeared every time.
PROTOCOL · KILL CRITERIA
Research note
The acceptance rules are written down before any test runs, so there is nothing to argue about after. Most candidates fail. Several ideas we liked failed too.
REJECTED · MICROSTRUCTURE SIGNALS
Research note
Before we trust a signal, it has to beat a shuffled version of itself. Most of the ones we tested couldn't.
REJECTED · PROTECTIVE EXITS
Research note
Profit-taking, loss caps, trailing exits — each one we tested removed the same handful of outsized winners that carried the whole backtest. None survived out-of-sample testing.
THESIS · NEGATIVE RESULTS
Research note
Most ideas we test fail. The failures are the only part of the record that isn't survivorship-biased, so nothing gets deleted — including the ones that were embarrassing.
Earlier writing — Medium, 2024–25
stock-market · finance · openai
Medium
A 2024 experiment: give a bot access to its own error logs and let OpenAI's o1 rewrite the parts that failed.
machine-learning · coding · data-science
Medium
The 2012 network that dragged computer vision into the deep learning era — and started the scale race the field is still running.
