Nonzero Lab Research · Method · 2026-08-14 · 6 min read

Can You Tell a Real Chart From a Random Walk?

Coin-flip charts fool chart readers into seeing patterns — yet with feedback, people learn to tell real series from random ones. Pattern skill is trainable.

The coin-flip chart

Malkiel (1973) records a famous anecdote: he had students generate fake price charts by flipping a coin — heads for up, tails for down — and then showed the drawings to a chart analyst. The analyst, not knowing where the chart came from, read familiar patterns into it and offered trading calls.²

The story is usually quoted to prove that chart reading is astrology. What it actually proves is narrower: humans see patterns in noise by default. It says nothing about the reverse question — between a real price series and a randomized one, can people learn to tell the difference?

Fairness also requires noting what kind of evidence this is: an anecdote, not a controlled experiment. It demonstrates that an analyst will read patterns into pure noise. It cannot demonstrate that all patterns are uninformative. That second question needs a stricter design.

A Turing test for markets, decades later

Hasanhodzic (2010) turned that question into a scoreable experiment. In a web-based game, participants were shown two return series at a time — one taken from real market data, one produced by randomly shuffling the real series — and asked which was which. The crucial design element was feedback: after each answer, you immediately learned whether you were right. Under those conditions, participants' ability to separate real from randomized improved steadily with practice.¹

The shuffling step is the elegant part of the design. Shuffling preserves the overall distribution of returns — the fatness of the tails, the frequency of extremes — while destroying the order in which returns arrived. If people can learn to tell the difference, what their eyes are picking up is not "how big the moves were" but how the moves were organized in time: the clustering of volatility, the rhythm of continuation and reversal. Those are exactly the things real markets have and pure randomness lacks.

Read together with the coin-flip story, the picture becomes coherent. Without feedback, pattern recognition produces hallucinations. With feedback, the same pattern recognition can be calibrated. The question was never whether people see patterns — it is whether the patterns you see have ever been corrected by an answer key.

Do patterns actually carry information?

A more rigorous layer of evidence is computational. Lo (2000) started from mathematical definitions of classic patterns — head and shoulders, double bottoms and the like — used kernel regression to detect them automatically across long histories of US stock data, and then ran statistical tests. The conclusion is specific and restrained: for some patterns, the conditional distribution of returns after the pattern differs detectably from the unconditional distribution.³

Two things about that design matter more than any single finding. The patterns were defined by formula, not eyeballed, so there was no discretion in what "counted" as a head and shoulders. And the test compared distributions, not stories — no room for a narrative about what the pattern "meant." This is what checking looks like when it is done carefully.

Note the boundaries of that sentence. It says "some patterns" and "the conditional distribution differs." It does not say any pattern predicts direction reliably, and it certainly does not say trading on patterns is profitable. What it supports is a weaker claim that matters more to a practitioner: price series are not pure noise, and at least some of the shapes the eye cares about carry statistical information.

"The conditional distribution differs" deserves a plain-language restatement. Pick a random day, and tomorrow's return follows one distribution; pick a day where a given pattern just completed, and tomorrow follows another distribution — and the two can be told apart statistically. A difference is not a profit: the gap may be tiny, transaction costs may eat it, another market may not show it. But it is enough to falsify the strong claim that charts contain nothing at all.

What actually teaches the eye?

String the three results together. People see patterns in noise by default (the coin charts). Some patterns genuinely carry information (the kernel tests). And the ability to tell real from random improves only with feedback (the Turing test). The conclusion follows: the bottleneck in chart reading is not how much you have watched, but how much of what you watched came with an answer key.

There is a quieter consequence hiding in that conclusion: measurement is the prerequisite of training. If you have never scored your eye against a known answer, you do not have an opinion about your chart reading — you have a mood about it. The first blind test you run is less a training session than a baseline measurement, and baselines are what make every later session interpretable.

This also explains why the gap between a ten-year screen-watcher and a beginner is not necessarily "experience." Watching without feedback is familiarity; judging with feedback is training. It is the same problem described in Why Your Chart Review Lies to You: when feedback is delayed or contaminated by narrative, screen time does not compound into skill.

A blind session is essentially this Turing test run on yourself: no ticker, no date, only the bars, a recorded decision each time, and then the reveal. Every reveal is one round of calibration.

The direction of calibration is worth remembering. The goal is not to become someone who can always tell real from random; it is to know how well you can tell. A person who knows their discrimination is at fifty percent makes far better decisions than one who believes eighty and is also at fifty — the former sizes confidence and risk to their true level, the latter does not. Knowing your own hit rate is not a consolation prize; it is the piece of information everything else hangs on.