Before asking whether your strategy predicts direction, ask whether direction is even worth predicting at your horizon.
The toll you pay is fixed: you cross the spread once, whatever happens next. The move you are paid for is not: it grows roughly as the square root of the holding time. So there is always a horizon below which even a perfect direction predictor loses money — and the only honest way to find it is to measure it.
HorizonProbe is a single MQL5 script that measures it, horizon by horizon, on your broker's real data and your broker's real spread.
It never trades. No orders, no positions, no account access, no
#include. Copy the one file into any terminal and run it.
For every horizon you list (in minutes), over every sliding window in your history:
| output | what it is |
|---|---|
| Ceiling (median cost) | share of windows where |close(t+H) − close(t)| exceeds the median round-trip cost |
| Ceiling (per-window cost) | the same count, but each window judged at the spread actually observed at its entry |
| ★ Oracle expectation | mean(|d|) − cost — this is the profitability judge |
| Break-even rate | 0.5 + cost / (2 × mean|d|) — the directional accuracy you need just to cover costs |
| Windows kept | and the non-overlapping equivalent (≈ used/H), because sliding windows are not independent draws |
A constant cost flatters you. Gold's spread explodes exactly when the move is large, so judging a news minute at the toll of a quiet hour overstates the ceiling. Printing both numbers turns that bias from an assumption into a figure you can read.
The ceiling counts windows that individually clear the toll. The expectation is mean|d| − cost.
An oracle with perfect direction earns (|d| − cost) on every window, so it is profitable if and only if mean(|d|) > cost. With fat tails — and gold's tails are fat, median ≪ mean — an oracle can be strongly profitable while fewer than half the windows clear the toll. Reading the ceiling as the verdict is systematically pessimistic, precisely on the instruments where this tool is most useful.
The script says so in its own output, because getting this backwards costs you a strategy you should have kept.
The usual shortcut is √t scaling plus a Gaussian. It is wrong in the direction that matters: a Gaussian under-states the large moves, which are exactly the ones that pay for the toll. HorizonProbe does no scaling and assumes no distribution — it walks every sliding window and counts.
★ A real result from using it. On XAUUSD, an estimate placed the profitability flip at around 30 minutes. The measurement put it at 5 minutes — a six-fold difference, and enough to reverse a verdict about an entire family of short-horizon strategies. Estimating that number and measuring it are not the same activity.
For scale, on the same instrument at M15: median bar amplitude ≈ 140 points, median spread 14–20 points. The toll is 10–14 % of a typical move. That ratio is what this script tracks across horizons.
- File → Open Data Folder in MetaTrader 5, then drop
HorizonProbe.mq5intoMQL5/Scripts/HorizonProbe/. - Open MetaEditor (F4) and compile with F7. Expect
0 errors, 0 warnings. ⚠️ Tools → Options → Charts → Max bars in chart → Unlimited, then holdHomeon the chart until it stops scrolling back. The script reads what the terminal has cached; the cap truncates silently and raises nothing.- Navigator (Ctrl+N) → Scripts → HorizonProbe, drop it on any chart of the symbol. It reads M1 regardless of the chart period.
git clone https://github.com/Sjrazaviebra/HorizonProbe.git| input | default | meaning |
|---|---|---|
HP_Horizons |
1,2,5,10,15,30,60,120 |
horizons in minutes, comma-separated, integers only |
HP_MaxBars |
500000 |
upper bound on the M1 history read |
HP_DateFrom / HP_DateTo |
1970.01.01 |
measured period; 1970 means all history read |
HP_Lots |
0.01 |
display only — converts points to currency, changes no ratio |
Results go to the Experts tab.
★ It refuses loudly rather than measuring quietly. A malformed or duplicated horizon, insufficient history, a truncated read — each stops the run with a stated reason. A silently truncated measurement is worse than no measurement.
- Sliding windows overlap (they share H−1 minutes) and are not independent draws. The non-overlapping equivalent (≈ used/H) is published, and it is that number which carries any sample-size rule.
- Strict counting: a move exactly equal to the cost does not pay the toll and is not counted.
MARGE = CEILING − BREAK-EVENis printed for reference, but⚠️ it subtracts two rates of different natures — a share of windows minus a required accuracy. Read it as an indicator, never as the verdict. The verdict is the expectation.- Gaps are reported as measured durations (
≥ 30 min/< 30 min). The script does not claim to tell a weekend from a data hole — it reports what it saw.
- Reads M1 and aggregates from it; horizons are expressed in minutes.
- Measures the market and its cost, not a strategy. A viable horizon is a necessary condition, never a sufficient one.
- Cost is the round-trip spread. Commission and swap, if your account has them, are yours to add.
Source-available for reference, evaluation and demonstration. See LICENSE.
Built by Javad Razavi — The Solution Maker · javadrazavi.fr
See also DataForge — the same idea applied to the data itself: know what is really inside your bars before you train on them.