Overfitting (Curve Fitting)
Picture a student who memorized the answer key to past exams. Retake those exact exams and they score 100 — but change the questions slightly and they flunk. An overfit trading rule is exactly this. It fits the 'past exam' of historical charts perfectly and falls apart on the 'new exam' of the future.
Here's how it happens in numbers. Say the rule 'buy when RSI drops below 30' fired 500 times historically with a 52% win rate. Not satisfied, you bolt on conditions — 'only on Wednesdays,' 'only in the early-morning hours,' 'only when RSI is below exactly 28.3.' Eventually you've built a rule that went 5 for 5 in past cases — a 100% win rate. But you haven't discovered a law of the market; you've reverse-engineered conditions to wrap around five past moments that happened to go up.
Why does this happen? Because if you try hundreds or thousands of combinations of conditions and numbers, a few of them are bound to fit the past by pure chance. Have 1,000 people flip coins and someone lands 10 heads in a row. That person isn't a coin-flipping master — and among thousands of tested combinations, the rule with the best historical score is likely the champion of luck, not skill.
There are telltale signs: the sample is small (fewer than a few dozen cases); the conditions are oddly specific (no one can explain why 28.3); tiny tweaks swing the results wildly (a goldmine at RSI 28, a wreck at 30); it only works in one particular period. If any of these applies, be suspicious.
The remedies aren't glamorous: keep rules simple, gather enough cases, split the data into periods and check the rule holds roughly everywhere, and always include fees. The rules that last tend to be the ones you can explain in a single sentence.
What the data actually shows
Barobara uses two safeguards against this trap. First, combinations with fewer than 40 historical cases are never published — the smaller the sample, the easier it is for luck to look like skill. Second, we don't hide the fact that shrinking the take-profit target can make a win rate look like 90%. On each signal page in the setup catalog you can see the whole curve — win rate rising while the expected P&L flips negative. Next time someone leads with a high win rate, picture that curve.Common misconceptions
'More conditions and more precision make a better strategy' — it's the opposite. Every added condition raises the risk of fitting only the past. Rules that survive tend to be simple.
'A 90% historical win rate means it's proven' — first ask how many cases. 9 out of 10 is well within the range of luck. And since win rate can be inflated just by tightening the take-profit target, a win rate alone proves nothing.
FAQ
Q. How do I check whether my strategy is overfit?
There's no perfect test, but check three things: is the sample large enough (at least a few dozen cases), is the rule simple enough to explain in one sentence, and do results hold up when you slightly change the rule's numbers? If any answer is no, suspect a rule that only fits the past.
Q. Are the flashy profit screenshots from paid signal groups overfitting too?
They may be rules curve-fit to past charts, or cherry-picking — showing only the trades that worked. There's one test that matters: do they publish the complete record, wins and losses, together with the conditions? A track record that doesn't show the full distribution isn't a track record.