OptiNod Academy
Recency bias — Mistaking a normal losing streak for a broken system
Even a tested positive-expectancy system can have long losing streaks. Compare them with expected streak length and track system changes before concluding the strategy failed.
A tested system with positive expectancy can normally produce a long losing streak. Remember only the last five or six trades and normal variation may look like a broken system.
Recency bias is the tendency to give recent outcomes more weight than they deserve. It is related to the availability heuristic described by Amos Tversky and Daniel Kahneman in 1973. Instead of counting how often something happened, people judge its probability by how readily examples come to mind. Five losses just experienced are vivid; 20 wins from six months ago are faint. The recent sample then dominates judgment.
For traders, this bias works in two directions. After several losses in a row, they declare “the system is broken” and stop trading. After a few recent wins from a different parameter setting, they decide “this setting fits the current market” and switch. Both treat a handful of trades as a reliable measure of the system's present state, though the sample is far too small.
Even a positive-expectancy system normally generates long losing streaks. For a system with win probability p repeated N times, the expected longest losing run is approximately log base 1/q of (N·p), where q is the loss probability (1−p). With a 45% win rate over 200 trades, the estimate is about eight. Eight losses in a row need not mean the system broke. With enough trades, such a run is a normal occurrence. Remembering only the latest five or six can make normal variation look like failure.
Availability gives recent outcomes too much weight
Recency bias is fundamentally about what memory makes easy to recall. Estimating a probability calls for the full history, but human judgment substitutes vivid examples. A fresh loss is recorded with emotion; details of older wins fade. The last five trades become the reference point while the previous hundred are left out.
The small-sample illusion compounds this. Five or ten trades are nowhere near enough to estimate a system's win rate, yet people expect a small sample to show a stable proportion. Tversky and Kahneman called this the law of small numbers in 1971: the mistaken expectation that a small sample will faithfully reflect the population. A 55%-win-rate system can readily experience four losses within five trades. If that short run is mistaken for its current win rate, a 55% system may be perceived as a 20% one.
Once a recent sample is overweighted, normal variation looks abnormal.
Positive-expectancy systems also produce losing streaks
Calculation, not emotion, tells you whether a streak is within the expected range. For N independent trades with win probability p and loss probability q = 1−p, expected longest losing-run length is approximately log base 1/q of (N·p).
Consider a trend-following system with a 45% win rate. Across 200 trades its expected longest losing streak is about eight. At 50% over 100 trades, it is about six. A system winning 40% of 150 trades can normally experience a run of roughly eight or nine losses. The more often a system loses relative to winning, the longer its normal streaks.
Frequency can also be estimated. Across 100 trades from a 45%-win-rate system, stretches of five losses occur more than twice on average. Even a 50%-win-rate system produces one or two such stretches on average. Five losses in a row are not an exceptional accident; they are a normal part of an accumulating sample.
The conclusion is straightforward: with a system's win rate and number of trades, you can estimate how long its worst ordinary streak may be. A streak within that range does not prove the system is broken.
Declaring normal losses a failure can switch the system off before a major trend
The error can be costly in real markets. From June through October 2024, Bitcoin spent roughly five months moving within a broad range from about $49,000 to $74,000. On August 5 it fell toward $49,000 before recovering into the $60,000s. Direction remained unsettled through much of that sideways period.
Trend-following and breakout systems tend to accumulate small losses in such a market. They enter a breakout that reverses, enter again and see another reversal. After months of these losses, recency bias makes “this system no longer works” feel plausible.
A trader who switched it off in October missed the next phase. Bitcoin began rising near $69,000 in early November and reached $108,353 on December 17, a gain of roughly 57%. Enduring small losses during the range was the cost of remaining available for a trend like this; recency bias can lead a trader to stop just before it appears.
Trend-following expectancy often depends on a few large moves. Small range-bound losses are the cost of waiting for them. Turn the system off during that run and you may pay the cost without capturing the trend.
Two numbers reveal recency bias
Treating recency bias only as an emotion makes diagnosis vague. Two figures show whether it is influencing decisions.
First, compare expected losing-run length with the actual run. Put the system's win rate and planned number of trades into the formula above. A run within the expected range is ordinary variation; a run well beyond it warrants a review. For a 45%-win-rate system across 200 trades, six losses remain below the estimate of about eight, so six alone are weak evidence for switching it off.
Second, count how often the system is changed. In the trade record, note each pause or parameter change and the losing-run length immediately before it. If changes cluster just after short runs still within the expected range, that is quantitative evidence of retiring the system too early. Compare performance after each change with a counterfactual in which the prior system continued to see whether the change helped or hurt.
- Estimate expected streaks: Calculate and record the expected longest losing run from the system's win rate and expected trade count before trading.
- Mark streaks at changes: Whenever you stop trading or change parameters, record the preceding losing-run length.
- Compare before and after: Calculate performance after a change alongside performance the previous system would have produced if maintained.
- Set a normal-range reference: Use expected streak length as a reference before abandoning the system; keep it running through runs inside that range unless other evidence changes.
Know the normal range beforehand to hold through ordinary streaks
Streaks unsettle judgment when the normal range of variation is unknown. Estimate it in advance and a run within that range is less reason to change course. OptiNod's validation tools can help estimate that range before trading.
Walk-forward analysis rolls through separate optimization and validation windows and shows how performance changes in each. The distribution across validation windows reveals how poor an ordinary stretch may look. Three robustness measures and period-by-period, such as quarterly, measurements split one system's performance across conditions. If one weak quarter is offset by others, that weakness may be within normal variation.
Monte Carlo resampling of trade outcomes serves a similar purpose. Reorder the same trades thousands of times and you obtain a distribution of plausible longest losing streaks and maximum drawdowns. Knowing where the current drawdown sits in that distribution helps distinguish a normal outcome from a warning.
With the expected range written down, the last five losses alone cannot overturn the plan.
Two pitfalls
Switching to parameters that performed best recently. The other side of recency bias is chasing a setting that just worked. Selecting last month's most profitable parameters likely fits that recent period too closely, a form of overfitting. Changing settings in response to recent returns effectively fits a curve to those returns, and the new setting can struggle in the next period.
Adding a rule only during a losing stretch. A trader may add “one more filter would have avoided these losses” during a streak. The filter is tailored to the losses just observed, which may be its entire evidence base. It can remove those losses from history while lowering expectancy across the full period. Always test a new filter on the full backtest and on independent data.
The normal range of losing streaks comes from calculation
Recency bias often acts most strongly when the system is still sound. A tested positive-expectancy system can normally produce long losing runs, which look like breakdowns when judged only by the most recent trades. The remedy is not a stronger resolve. Before starting, estimate the expected longest run from win rate and trade count and set a reference for when a review is warranted. If the present streak is within that range, the last five losses alone do not justify shutting the system off. Memory cannot tell normal variation from failure; calculation can.