The Edge That Wasn't
A Memoir of Retail FX, and What the Receipts Showed

The Edge That Wasn't

Algorithmic Trading, 81 Experiments, 4,932 Live Trades, and the Answer Was No

Front cover of The Edge That Wasn't

Most books with "algorithmic trading" on the cover sell a system that prints money. This one hands you the losses. The central artifact of the book is a table of broker-verified results — its own — and the frozen snapshot of every recorded trade ships with it, so any per-strategy row reproduces from the raw file.

−55,000 pips of bad signals from one lookahead bug, across sixteen strategies
~$200 funded across thirteen broker sub-accounts, peaking near $199
+171 pips lifetime, the best of thirteen accounts — later wound down like the rest
4 headline results retracted, corrections published

The trade ledger — every closed trade, 5,556 rows — is committed here: trades_snapshot_2026-05-31.csv. The raw S5 price history is not redistributable; the analysis scripts pull it from the OANDA v20 API. See Verify for what's in the file and how to query it.

It is the record of fx-core, a containerized retail FX research project. No single strategy ever ran on more than about a hundred dollars, and most on far less; roughly two hundred dollars was funded across the thirteen broker sub-accounts over the project's life, peaking near a hundred and ninety-nine. The methodology is the subject, not the size of the account.

One question runs through it: is there a repeatable, statistically real edge available to a retail trader in spot FX, and if so, is it large enough — after costs, after tail risk, after the capital required to hold open positions through their drawdowns — to beat a stock-index fund? Eighty-one audited experiments answer it: thirteen positive on their own terms, twenty-two mixed, forty-five negative, and one late survivor retracted.

The apparatus included a positive control — an instrument calibrated to prove it could detect a real edge before being trusted to report that there wasn't one — and it forced four published retractions on this book's own results. Both are documented here: the calibration and the corrections.

What the apparatus taught outlasts the trading. Pre-registration, a sealed out-of-sample set touched exactly once, a working taxonomy of lookahead bias, the coin-flip control that separates skill from luck, and the cost floors that sink most edges: a full research-methods spine, demonstrated on live results rather than lectured. It is meant to be useful to a data scientist or engineer who will never place a trade.

Front cover of The Edge That Wasn't

Get the book — Paperback, 582 pages, $29.99

Buy on Amazon → Read an excerpt

Live on Amazon now (allow up to 3 days for the in-stock flag in some marketplaces). Kindle edition in preparation.

Every number is checkable · star the repository to catch the release

Inside

"The positives are not victories in the usual sense. Several did not survive finite-margin re-validation once a realistic stop-loss was added; two worked in simulation but failed live execution; and the late-program successes are cost-accounting and risk-reduction findings, not edges."

— from the repository's experiment map. Four headline results were later retracted outright — the corrections are published here.
Get the book → or read an excerpt from Chapter III first

Occasional updates only — corrections and any post-book result that survives the gates: star the repository to follow, or open an issue there to reach the author.