The Signals page publishes a direction and a strength for ten currency pairs every day, from three daily analyzers: DailyTrend, Breakout and DailySmaSlope. We had never measured whether those readings predict anything. On 9 September 2026 we did, and the answer changes how the page should be read.
What we measured
Each analyzer is a pure function of closed daily bars, so its output over any history can be rebuilt exactly from the rules in the published source (Signals/D1_Signal_*.mq5 in the public repository). We rebuilt all three over ten years of daily bars, September 2016 to September 2026, on all ten pairs, and for every day a signal was on we recorded the direction-signed return over the next 1, 3, 5 and 10 days.
If the signals had predictive value, those forward returns would be positive.
What we found
They are negative. On every strategy, at every horizon.
| Strategy | Signal-days | 5-day mean | 10-day mean | Hit rate (5d) |
|---|---|---|---|---|
| DailyTrend | 10,384 | −5 bp | −10 bp | 48% |
| Breakout | 1,361 | −16 bp | −23 bp | 45% |
| Breakout, strength ≥ 80 | 99 | −101 bp | −129 bp | 34% |
| DailySmaSlope | 19,782 | −4 bp | −7 bp | 48% |
The t-statistics run from −3 to −6. The windows overlap, so the magnitudes are optimistic, but the sign is not in doubt. And there is a gradient: the stronger the reading, the worse the forward return. A Breakout reading of 80 or more — a close above the 20-day range with expanding volatility — was followed by an average move of about one percent against the signal over the next five days, and went the signal's way only a third of the time.
We also ran each signal as a naive strategy: hold the direction while the signal is on, go flat when it reads FLAT, ten pairs equal-weight, one pip per flip. All three lost money over the decade.
Why this is not surprising in hindsight
Every intraday result we published the same week points the same way. The one system that survived out-of-sample testing is mean reversion in the European morning; three trend-following entries were rejected; fading breaks in the New York overlap lost, and so did riding them. The majors, on the timeframes and years we can test, revert. A strong daily trend reading is therefore not a forecast that the move continues — it is a measurement of how far it has already gone.
Does the inverse work?
Partly. Fading a Breakout reading of strength 80 or more, one event per pair per five days, gave about +100 bp per event on a five-day hold in both halves of the decade, on nine of ten pairs. That was strong enough to build an engine and test it on our broker's own daily bars from 2013. On that data the sign replicated and the size did not: roughly nine pips a trade at 0.01 lots, flat before 2020, profit factor 1.12–1.14. It sits on the presets page as a candidate that did not pass our promotion gate, with its numbers. We are not trading it.
How to read the Signals page now
The analyzers have not changed, because the readings are informative — just not in the way the labels suggested. Read a strong LONG as "this pair is extended to the upside", and a strong SHORT as the mirror. That is a useful thing to know about a market. Read as "this pair will go up", the same reading has been wrong more often than right for ten years.
The page says this now. The full study — method, tables, the replication and its caveats — is in the working notes that accompany the public repository, and the script that rebuilds the signals is there for anyone who wants to check the arithmetic.
Nothing on this site is investment advice. The readings are measurements of price behaviour, published so that they can be checked.
Comments