After a 16-Year Audit, We Retired Our Signal Logic — Here Is the Transition Plan

On 21 July 2026 we halted new entries from the old logic behind our public FX signal service (an H4 portfolio of 17 configurations). The reason is simple: a 16-year audit proved the design cannot win.

This post is the announcement of that retirement and of the transition plan to a new logic. Every number below is a measured result. Our past track record and third-party verification (Myfxbook) stay untouched — nothing gets deleted or reset.

What happened

• Since the public track-record epoch (13 July): 3 wins, 3 losses, −114.4 pips total. Average win +7.4 pips; average loss −45.5 pips.
• A 16.1-year out-of-sample re-verification (4,947 trades): 73.5% win rate, −4,670 pips.

A system that wins 73.5% of the time and still loses money is not unlucky. It is working as designed.

Why a 73.5% win rate loses

Place a bracket with target distance a and stop distance b on a random walk and the target gets hit first with probability b/(a+b) (the gambler's-ruin result). Our old design (target 0.5–0.8×ATR, stop 1.5–2.0×ATR) implies a theoretical 74–76% win rate. We observed 73.5%.

The decisive test: across every bar of 16 years × 21 pairs, entering in either direction with no signal at all, a 10-pip target against a 2×ATR stop wins 83.7% of the time (theory: 83.3%). Buy win rate plus sell win rate sums to 167% — two times out of three, both directions "win".

The win rate was never evidence of skill. It is the output of the target-to-stop distance ratio. And the unconditional 83.7% sits below the breakeven requirement of 86.7%; the gap compounded for 16 years as spread cost, −0.94 pips per trade.

Exits cannot fix this — the parameter surface has a train-to-test rank correlation of ρ≈0. We searched 252 time-of-day cells for a pocket that beats breakeven in both halves of the data: zero survived. We also tested a machine-learning filter with a pre-registered acceptance gate (AUC ≥ 0.55); it scored 0.51–0.52 out-of-fold and was rejected as declared.

The full methodology is in the companion piece, "My Trading System Has a 73.5% Win Rate and Loses Money". Reproduction code and result data are public on GitHub.

The new signals and the timeline

What survived 16 years of falsification was not a price pattern but an information event:

• Post-announcement drift: when a major economic release misses consensus by a large, volatility-normalised margin, price keeps adjusting in that direction for roughly 24 hours. Measured: +10,143 pips over 15 years (8 pairs, spread-deducted); only USD major pairs survive a 2× cost stress test.
• The structure is the inverse of the old design: wide stop (3×ATR), profit left to the drift, so the average win exceeds the average loss. The "one loss erases nine wins" problem disappears at the root.

We are not flipping it on immediately. It currently runs as a shadow (no orders placed), tracking real releases in real time and recording how live spreads and execution prices compare with the backtest's assumptions.

Transition timeline (planned):

1. Now through August: shadow accumulation across US releases (NFP, CPI, retail sales, GDP, weekly claims — an estimated 15–25 signals per month)
2. Late August: at ~30 measured signals, we compare shadow results against the 15-year out-of-sample expectations
3. If it passes, new signals go live in early September, with a reformatted delivery: results reported in R multiples (profit per unit of risk) instead of raw win rates

To be explicit: if it does not pass, it does not ship. We no longer sell good-looking win rates. We publish the expectancy that survives falsification, and nothing else. During the transition there will be no new signals; progress updates will be posted on Discord.

Why publish this at all

The standard move in the signal industry is to delete a losing track record and start a fresh account. We are doing the opposite: publishing the verdict, keeping the history, and resetting nothing.

Sixteen years, half a million bars and 4,947 trades taught us one thing worth repeating: there is no holy grail — only small edges that survive falsification, and the discipline not to break them. That discipline, with everything published, is the product.