OptiNod Academy
Confirmation bias — More supporting indicators can strengthen conviction without improving direction
Gathering more chart evidence does not necessarily improve a forecast. Compare in-sample with out-of-sample expectancy and count whether extra reasons improve realized R.
Gathering more reasons from a chart does not necessarily make your directional call more accurate. What often grows is only conviction in a direction already chosen.
Psychologist Peter Wason demonstrated confirmation bias in a rule-discovery experiment in 1960. Participants saw the sequence 2, 4, 6 and had to discover the hidden rule. Most formed a guess such as “increase by two,” then offered confirming sequences like 8, 10, 12. The actual rule was simply “numbers in ascending order.” Those who tested sequences that could disprove their own guess were the ones who found it. In short, people look for information supporting a judgment already made and rarely seek information that could refute it.
Popular advice treats the problem as an attitude: “be objective” or “leave emotions out.” Confirmation bias does not arise from too little analysis; it operates during analysis. The more evidence you seek after choosing a view, the more supporting evidence you tend to notice. “Analyze more” can therefore make it worse. A better attitude alone is insufficient. You need a procedure that looks for disproof before the judgment hardens.
The cost is not measured by the count of entry reasons. Decide to go long and the chart can offer endless support: oversold RSI, Fibonacci support, bullish divergence and a demand zone may all appear on one screen. Opposing clues on that screen were simply not counted. If five reasons do not improve the directional edge, the extra reasons added conviction, not information. This resembles fitting extra backtest conditions to the past. The effect can be measured by the gap between in-sample and out-of-sample expectancy.
Searching longer can leave direction unchanged while conviction hardens
The order in which information is gathered is central. People choose a view, then seek information that supports it. Before choosing direction, the same chart may show similar numbers of bullish and bearish clues. Once a long is chosen, bullish clues stand out and bearish ones may not be sought at all.
Charts make that sequence dangerous because potential reasons are plentiful. One screen can contain dozens of indicators and hundreds of bars; choose either direction and some combination will support it. RSI looks oversold, a prior low aligns with Fibonacci support, or volume has risen. Confidence grows with each supporting clue, but the directional edge has not changed if contrary clues on the same chart were never counted.
The 2–4–6 experiment illustrates the point. Any number of sequences consistent with your guess do not prove the guess. You need to test a sequence that could refute it. Entry arguments work the same way: five supporting clues do not prove direction. Evidence gains weight when the conditions that would disprove the view are also tested.
Adding conditions to strengthen conviction resembles curve fitting
Adding supporting reasons to every entry parallels adding filters to a backtest to fit the historical curve. Both add conditions that favor a conclusion already chosen. In a backtest, “this filter removes past losses” becomes another rule. On a chart, “this indicator is bullish too” becomes another reason. More conditions explain the past increasingly well.
The problem is that additional conditions may fit the past alone. Add a filter and in-sample backtest performance will often improve. If it happened to exclude a few historical losses by chance, the effect need not appear out of sample. Likewise, if entries supported by five arguments perform no better than those supported by two, the extra three increased conviction without adding an edge.
Multiple testing explains the illusion. Try many candidate conditions and some will improve historical results purely by chance. Searching broadly for supporting indicators is similar. Look across enough measures and a few will certainly point bullish; gathering them can feel like independent confirmation. Curve fitting inflates in-sample performance, while one-sided evidence gathering inflates entry confidence.
Near a high, supporting reasons are easy to find
Confirmation bias can be most costly near the end of a trend. Bitcoin reached about $126,200 and a then-record high on October 6, 2025. A trader committed to a long lacked no supporting reasons: upward momentum making new highs and closes above a rising trendline could both be cited. Contrary clues were on the same chart but went uncounted.
Price then fell roughly 36% to $80,600 by November 21. The place where bullish reasons were easiest to see was also highly risky. The longer an advance has run, the more bullish evidence describes the rise that has already happened, without predicting the next direction. Confirmation bias makes opposing evidence particularly easy to miss then.
A similar pattern appeared in early 2026. Price was $97,924 on January 14 and fell about 39% to $60,000 by February 6. Plenty of reasons near the high could support a chosen long, but the decline that followed over a little more than three weeks occurred regardless. The number of supporting reasons is not a measure of directional edge.
Two numbers can measure confirmation bias
If it is treated only as an attitude problem, there is no clear place to intervene. Pair each entry's reason count with its realized R and calculate two figures.
First, the gap between in-sample and out-of-sample expectancy. Split historical data into an earlier period for fitting conditions and a later period for testing the unchanged strategy. Adding supporting conditions can raise in-sample expectancy while out-of-sample expectancy stays low. The gap estimates the portion inflated by fitting the chosen view. Walk-forward analysis repeats the comparison across rolling windows to see whether the gap persists.
Second, the relationship between the number of reasons per entry and realized outcome. Record how many supporting reasons existed at entry, then pair that number with realized R. If entries with many reasons have similar average R to those with few, the added reasons provided little information. Robustness checks also flag strategies whose performance collapses after small changes in conditions; a rule with that sensitivity may merely fit history.
Write down what would disprove the idea first
Confirmation bias starts working after direction is chosen. To counter it, examine opposing evidence before committing to direction. Before entering, write what would show that your view is wrong. That reverses the usual order of collecting only support.
- Record the opposing scenario first: Before confirming entry, write the condition that would invalidate your direction. Count supporting clues afterward.
- Put invalidation into an order: Submit a stop at the invalidation level when entering. Do not leave contrary evidence as a judgment to reconsider later.
- Record the reason count: For every entry, note supporting reasons and periodically compare the count with realized R.
- Test out of sample: If a new condition raises conviction, run the strategy fitted in sample unchanged on out-of-sample data and compare expectancy.
- Reserve a test interval before judging: Set aside out-of-sample and walk-forward periods before using live or paper returns as the only test of whether a strategy “makes money.”
The aim is not to remove confidence. It is to make validation results, rather than a tally of agreeing clues, the basis of that confidence.
Two pitfalls
Believing you examined both sides. Simply looking at bullish and bearish clues does not remove the bias. People can accept supporting evidence and dismiss opposing evidence as an exception. If contrary evidence is seen but waved away, the outcome is the same. Turn it into a concrete invalidation order rather than leaving it to later judgment.
Assuming more reasons mean more certainty. Five overlapping reasons do not necessarily improve a directional call. Oversold RSI and oversold Stochastic, for example, may describe the same price decline in different forms. Counting overlapping signals repeatedly raises confidence without adding independent information.
The number of reasons does not measure directional edge
Confirmation bias can worsen with hard work. The more time spent looking after a direction is chosen, the more supporting clues appear and the stronger confidence becomes, without a corresponding improvement in edge. It is not caused by insufficient analysis, so “analyze more” is not the remedy; change the order of testing. Write the opposing scenario before settling on an entry and validate new conditions out of sample. Pair reason count with realized R, and measure the expectancy gap between in-sample and out-of-sample periods. Then “I have so many reasons” becomes a claim you can test.