Most machine learning models output a score. That score might be 0.87, or 0.23, or 0.61. But here's the problem: a score is not a probability. It looks like one. It sits between 0 and 1. But it doesn't behave like one — not without calibration.
Platt calibration (the technique that converts a model's raw output scores into actual, testable probability estimates) is one of the least-discussed concepts in algorithmic trading signals. That's a shame. Because if you're using a signal provider that publishes "73% win rate" without showing you how that number was derived, you're probably reading a fiction.
Here's how it actually works — and why we built it into Trade Forge's engine.
---
What a Raw Model Score Actually Means
Let's say a classification model outputs 0.85 for a long signal on EURUSD. Most people read that as "85% chance this trade wins." It isn't.
Raw scores reflect a model's confidence relative to its own training distribution. They're not calibrated against real-world outcomes. A model trained on imbalanced data (say, far more losing trades than winning ones) will systematically distort its scores. High scores might only deliver 60% win rates. Low scores might still win 40% of the time.
Without calibration, the numbers are internally consistent — but externally meaningless.
---
How Platt Calibration Fixes This
John Platt introduced the technique in 1999 to fix exactly this problem. The method is elegant in its simplicity.
You take the model's raw scores. You feed them into a logistic regression — a simple S-shaped curve — fitted against actual observed outcomes. The logistic function maps the raw scores to real probabilities that match historical base rates.
If your model says 0.85 on 1,000 signals, and 820 of them won, the calibrated output approaches 0.82. The number now means something. It corresponds to a frequency you can verify.
That's the entire point. Calibrated probabilities are honest probabilities.
---
Why This Matters for Forex Signal Providers
Here's a hot take: most signal providers hide their live results because the results are bad. Not because they're malicious — because their models aren't calibrated, their backtests don't generalise, and publishing real data would expose that immediately.
A calibrated model can be held accountable. You can plot its predicted probabilities against actual win rates across probability buckets (say, signals scored 0.6–0.7 vs. 0.7–0.8 vs. 0.8+). If the model is well-calibrated, the win rates track the predictions closely. This is called a reliability diagram, and it's one of the clearest tests of a signal model's honesty.
We publish ours. You can check it against our live trade log at tradingforge.io.
---
The Calibration Pipeline at Trade Forge
Our calibration engine isn't static. It retrains weekly on our full live signal archive — currently 2,400+ signals across forex, crypto, commodities, indices, and ETFs.
The process runs in three stages:
- Score generation. The base model outputs raw scores for each signal candidate across our 24 forex pairs and other asset classes.
- Calibration. A Platt logistic layer fits those scores against the prior 90-day outcome window. Session-specific and pair-specific adjustments are applied separately (in our data, signals fired during off-session hours carried a 14 percentage point lower win rate than on-session equivalents — that gets baked in).
- Threshold filtering. Only signals that clear a calibrated probability threshold get published. The threshold varies by signal style: scalp, day trade, and swing signals carry different base rates and different risk profiles.
This isn't manual discretion. It's a systematic, repeatable process with a documented audit trail.
---
Calibration vs. Raw Accuracy: A Direct Comparison
| Metric | Raw Model Output | Platt-Calibrated Output |
|---|---|---|
| Win rate accuracy | Overstated by 8–20% typically | Tracks observed rate within 2–3% |
| Reliability diagram fit | Poor — scores cluster at extremes | Smooth diagonal alignment |
| Threshold utility | Arbitrary | Statistically meaningful |
| Accountability | Hard to verify | Directly testable |
| Retraining requirement | Infrequent | Weekly minimum recommended |
The difference isn't academic. It changes which signals get published and which don't.
---
Who Doesn't Need to Know This
If you're a long-term investor — think monthly rebalancing, buy-and-hold ETFs, passive index exposure — calibration models are irrelevant to your workflow. You don't need algorithmic trading signals at all. This entire article is aimed at active traders who are evaluating signal models or building their own.
If you're using a discretionary strategy based on price action or fundamentals, calibration still doesn't directly apply. It's specific to model-based signal generation where a score is being produced and used as a decision input.
This concept matters most if you're: assessing a forex signal provider's methodology, building your own ML-based trading system, or trying to understand why your backtest win rate doesn't match your live win rate (R-multiples — how much you made relative to what you risked — are the other half of that story).
---
Frequently Asked Questions
Does Platt calibration improve win rates?
No — it improves the accuracy of predicted win rates. It makes a model more honest, not more profitable. A well-calibrated model tells you a signal has a 64% chance of winning. Whether 64% is good enough depends on your R-multiple and position sizing.
How is Platt calibration different from isotonic regression?
Both are post-hoc calibration methods. Platt calibration uses a parametric logistic curve — it's smoother and works better with smaller datasets. Isotonic regression is non-parametric and fits more tightly but can overfit with limited data. We use Platt because our weekly retraining windows are relatively short.
Can I see Trade Forge's calibration data?
Yes. Our methodology page and live signal log are published at tradingforge.io. We document predicted probabilities alongside outcomes so you can run your own reliability analysis.
Is this relevant for the best forex signals in 2026?
Calibration is becoming a baseline expectation for serious algorithmic trading signal providers. If a signal provider can't explain how their win rate numbers are generated — or won't show you live verified data — that's a meaningful red flag.
---
If you want to see calibrated signals in practice before committing to anything, our free Telegram and Discord channels carry the same signals as VIP (with a short delay). Watch how the signals perform live. Form your own view.
Join the free channel → tradingforge.io
---
This article is for educational purposes only. It does not constitute financial advice. Trading involves risk of loss.