Open Instagram and you’ll see it: “strategy with a 70% win rate”, “signal group with 85% accuracy”, “bot with a 90% win rate”. We decided to stop arguing opinions and start measuring. We ran 12 rounds of study on 45 days of real M1 candles, downloaded from Deriv’s official API, across 8 forex pairs — with strict separation between training and test data, plus a final window that no study had ever seen. The result is uncomfortable for anyone selling a course: the honest, stable ceiling came in at ~60-61% accuracy. The famous “70%” did show up — but only in cherry-picked slices that collapsed the moment they met new data. In this article we show the full numbers, the classic overfit that would have burned real money, and why an honest 61% is worth more to South African traders than a fantasy 70%.
The entire study was done with real data from Deriv — the broker with an official API and real forex 24/5.
Open a free Deriv demo account →How the study was done (full methodology)
Data: 45 days of REAL M1 candles from Deriv (official API), 8 forex pairs.
3-layer validation: 70% training / 30% test + a FRESH WINDOW — a block of data (12 June → 2 July) that no study saw during development.
Metric: accuracy on the direction of the next candle.
Break-even: with an 87% payout, you need 53.5% accuracy just to break even. Win rate ≠ profit.
Golden rule: a number that only shows up in one window is luck or overfit. Only what survives in all of them counts.
What exactly was tested
This was not a quick weekend test. Over 12 rounds of study, the same validation pipeline processed: 10 “normal” strategies (the ones every YouTube channel teaches), the same 10 strategies inverted (trading against the signal), time-of-day filters, an ADX filter, confluences of 2, 3 and 4 simultaneous signals and another 5 extra strategies. All with the same anti-overfit discipline: a strategy was only considered valid if its training performance repeated in the test set — and then in the fresh window, which acted as the experiment’s “real world”.
That last step is the one almost nobody takes. It’s easy to find a combination of indicators that “hit 70%” in the past: with enough parameters, you can always find a beautiful slice. What’s hard is for that combination to keep hitting on data it has never seen. That is exactly where the promises die.
The result: the honest ceiling is ~60-61% — and always on USD/JPY
After everything was tested, inverted, filtered and combined, the best stable performance — the kind that repeats in training, in the test set and in the fresh window — consistently landed in the range of 60% to 61% accuracy, always on USD/JPY, with mean-reversion strategies. No honest configuration went beyond that in a sustained way. None.
| Strategy family | Validated result | Verdict |
|---|---|---|
| Mean reversion (USD/JPY) | ~60-61%, stable across windows | The honest ceiling |
| “With the tide” gate | 66.9% in test → 51.2% in the fresh window | Classic overfit |
| Trend (MACD, SuperTrend, Ichimoku, pullback, strong trend) | 46-49% | Losers on M1 |
| Inverted trend | 52-54% | Weak edge |
Where the “70%” comes from: the anatomy of an overfit
The most instructive example in the study was a filter we called the “with the tide” gate — only trade when the signal agreed with the bigger move. On the test set, it nailed 66.9% accuracy. If we had stopped there, we’d have a viral post: “almost 67% proven on real data!”. But then came the fresh window — data no stage of the study had seen — and the very same filter crashed to 51.2%. Below break-even. A filter that looked like a gold mine would, in practice, have burned real money.
This is how the marketing “70%” is born: someone tests dozens of combinations, finds the slice that shone by chance, takes a screenshot and sells it. The slice doesn’t survive new data — but the buyer only finds that out with their account already in the red. If the seller doesn’t show out-of-sample validation, the number is worth nothing.
Trend strategies lose on M1 (and inverting them barely helps)
Another result that contradicts the conventional wisdom of the courses: every trend strategy we tested — MACD, SuperTrend, Ichimoku, pullback and “strong trend” — landed between 46% and 49% accuracy. In other words: losers, some of them losing worse than a coin flip. On the 1-minute chart, “follow the trend” simply did not work on our data.
Curiously, when we inverted those same strategies (entering against the signal), they moved up to 52-54% — a real edge, but a weak one, very close to the 53.5% break-even. That reinforces the reading that real forex on M1 has a dominant mean-reversion behaviour: overstretched moves tend to give back, and that is where the small statistical advantage that exists actually lives. We dig deeper into the comparison with the synthetic market in the article OTC is a random walk: the proof with data.
Indicator confluence: more signals is not more accuracy
We tested requiring 2-of-4, 3-of-4 and 4-of-4 reversal indicators to agree before entering. Intuition says more confirmation = more accuracy. The data says otherwise: reversal confluence added nothing. The reason is mathematical, not mystical: Bollinger, Z-Score and CCI are redundant — all three measure, with different formulas, the same thing: “price is at ±2 standard deviations from the mean”. Requiring all three to agree is asking for the same confirmation three times.
The only statistically independent signal in the group was the RSI, which measures the speed of the move rather than distance from the mean. Practical conclusion: the Double Reversal (Bollinger ∧ RSI) already IS the confluence — it combines the only two signals that bring different information. Stacking the rest on top is decoration.
The maths: why an honest 61% beats a fantasy 70%
With an 87% payout, break-even is 53.5% accuracy. Trading at a real 61%, the expected value is roughly +14% per trade — a concrete statistical advantage that, with serious risk management and trade volume, compounds over time. The “70%” from an overfit slice, on the other hand, is worth exactly zero, because with real money it turns into 51%, and 51% with an 87% payout is guaranteed loss in the long run. And no: martingale doesn’t fix a strategy without an edge — it only speeds up the blow-up.
Honest summary: anyone selling you “70-90% accuracy” is selling overfit or a lie. The ceiling that survived 12 rounds of study, train/test and a fresh window was ~60-61%, on USD/JPY, with mean reversion. It’s less glamorous — and it’s the only number you can build anything on.
FAQ — Win rates in binary options
Is there a proven strategy with a 70% win rate? In our 45 days of real data, no. Numbers above ~61% only appeared in slices that collapsed in the fresh window — the typical overfit pattern.
Is a 60% win rate profitable? Above 53.5% (break-even at an 87% payout), there is positive mathematical expectancy. At 61%, EV is ~+14% per trade. But win rate is not profit: without risk discipline, any edge dies.
Which pair had the best result? USD/JPY, consistently across all windows, with mean-reversion strategies.
How do I spot an overfit number? Always ask: was this result validated on data the strategy has never seen? If the answer is “no” or “I don’t know”, treat the number as advertising.
Want to replicate the study? Deriv offers an official API, historical data and a free demo account.
Create a Deriv account →Read next
→ OTC Is a Random Walk: The Proof With Data (and Where Real Edge Exists)
→ Best Hours and Days to Trade Forex M1: What the Real Data Shows (with SAST times)
→ Deriv Bot: How to Create Trading Bots Without Coding
→ Deriv Review 2026: Is It Safe for South African Traders?
→ All Broker Reviews for South Africa
About the author — Dan Machado
Founder of IA Trader Pro. Builds open-source trading bots and publishes studies with real data — no account screenshots, no promises of riches. Everything in this article came from code and data that anyone can reproduce.
Disclaimer: binary options and derivatives are extremely high-risk products and most retail traders lose money. This content is strictly educational and does not constitute investment advice, an offer or financial advice. The results presented come from a statistical study on historical data and do not guarantee future results. This article contains affiliate links. Always test on a demo account before risking real money, and never trade with money you cannot afford to lose.