Technical Analysis: The Magic That Isn't

A starling murmuration has no leader. No bird knows the plan, none sees the whole — each one reacts only to the six or seven neighbors around it. And yet the flock moves like a single organism: it turns, splits, evades, as if someone had given a command that nobody ever gave.

That is the most honest way into technical analysis I know. The lines on the chart cannot see the future. But hundreds of thousands of people look at the same lines and react to each other — and out of that emerges something that feels like prediction. The “magic” of TA is not magic. It is coordination. And that, as I want to show, is a far more interesting explanation than sorcery — because it also explains when the spell fails.

Illustrative image, AI-generated.

What technical analysis claims

Briefly, for anyone new: technical analysis reads patterns out of past prices and volume — support and resistance, trends, breakouts, indicators like RSI and MACD, levels like Fibonacci retracements — and infers likely next moves. The vocabulary itself lives in the trading terms index; this post is about the one question that had no room there: can this work at all — and if so, why?

What the research actually says

The short answer: it’s more complicated than either camp claims.

The most famous pro study is Brock, Lakonishok and LeBaron (1992): simple rules — moving averages, range breaks — showed statistically significant predictive power across ninety years of Dow Jones data. The result even survived the harshest test: in 1999, Sullivan, Timmermann and White bootstrapped across the entire universe of possible trading rules to check whether the finding was mere data snooping — it wasn’t. But the same study also found: in the more recent period (1987–1996), the edge had vanished. Remember that pattern; it returns shortly.

Lo, Mamaysky and Wang (2000) automated pattern recognition and found that classic chart patterns do carry statistical information — but “informative” is not “profitable,” and transaction costs weren’t part of the test. The big literature survey by Park and Irwin (2007) counted 56 positive, 20 negative and 19 mixed studies — while attesting most of them methodological trouble: rules picked after the fact, flattering cost assumptions, data snooping.

The opposing position deserves equal seriousness: under the efficient-market hypothesis (Fama, 1970), all information in past prices is already in the price — not because markets are dumb, but because many smart people price it in immediately. And the apophenia critique has a venerable exhibit: as early as 1959, Roberts showed that charts made from pure random numbers are visually near-indistinguishable from real price series. Malkiel tells the fitting anecdote in “A Random Walk Down Wall Street”: students generated a fictional chart by flipping coins, and a chartist saw a bullish formation in it and urged an immediate buy. (A book anecdote, not a study — but one that lands.)

And then there is the one big exception on which research largely agrees: momentum. What ran well over months tends, on average, to run a bit further — documented since Jegadeesh and Titman (1993), broadly replicated, and by now confirmed for crypto in top journals: as a time-series effect, as a cross-sectional factor, and most recently as a trend factor across 3,000 coins that survives transaction costs. Before that becomes euphoria, the footnotes: this is portfolio evidence on monthly horizons — no proof that chart-reading works in a five-minute window. And momentum has a documented price: rare, brutal crashes precisely in the panic and recovery phases where you’d least want them.

Illustrative image, AI-generated.

The trick: it works because everyone is watching

Now for the interesting part — the mechanism explaining why lines without predictive power can still have real effects. It is not speculation; it has been measured, most thoroughly by Carol Osler, then at the Federal Reserve Bank of New York.

Three findings that together explain half of TA:

First: when Osler analyzed support and resistance levels published by real FX desks (2000), the vast majority ended on round numbers — and the levels genuinely helped predict intraday trend interruptions. Second, and this is the most beautiful finding in the whole literature: using the first available data on actual customer orders, Osler (2003, Journal of Finance) showed that take-profit orders cluster right at round numbers — which is why trends reverse there — while stop-loss orders cluster just beyond them — which is why prices run unusually fast after a break. Two core predictions of charting, derived not from the lines but from the distribution of the orders themselves. Third: clustered stops can trigger each other in cascades (2005) — one layer of stops pushes the price into the next. What traders call a “stop run” has been academically dissected as a mechanism.

Add the anchoring effect of widely watched average lines (Avramov et al., 2021), and the question of whether enough participants even play along for coordination to hold. Answer: yes — in Menkhoff’s survey of 692 fund managers, 87 percent used technical analysis; on weekly horizons it outranked fundamental analysis. That is not retail folklore; those are professionals.

Fibonacci is the acid test for this theory — and passes it: the predictive evidence is thin to absent, the 50% level isn’t even a Fibonacci ratio, and yet the levels sometimes work. The most plausible documented explanation: they work because millions draw them and place orders there. George Soros coined the term reflexivity for this feedback — expectations altering the very market they claim to describe. A concept, not a study; but it is the right word.

And here sits the built-in self-destruction: coordination only holds while enough people play — and once a pattern is too well known, it gets traded away. McLean and Pontiff quantified this: published return anomalies lose an average of 58 percent of their power after publication. The flock turns — and whoever keeps flying the old direction flies alone.

In crypto, the loop is tighter

Everything so far applies to markets in general. Crypto adds an amplifier, and it has a precise reason: there are barely any fundamentals to hold on to instead. A remarkable study (Detzel et al., 2021) shows exactly that: for assets with hard-to-value fundamentals — Bitcoin is their prime example — technical analysis emerges rationally, because the price itself is the only shared source of information. Where no earnings per share exist, the 200-week line becomes the campfire everyone gathers around.

Then there’s the market’s structure. It is retail-dominated — the BIS analyzed crypto app data from 95 countries and found that about 40 percent of new users are men under 35, the most risk-tolerant segment (their back-of-envelope estimate: roughly three quarters of app users likely lost money on Bitcoin purchases through 2022). This crowd looks at identical tools: practically every exchange app ships the same preset indicators that the exchange academies themselves teach as the standard set — RSI, moving averages, MACD, Bollinger. (And since an exchange comes up, the obligatory line: that describes their teaching materials, not a recommendation; what sits on exchanges is yours on paper only.) The market trades 24/7, without halts, without a closing auction — nothing brakes a cascade once it runs.

And the forced flows are more visible here than anywhere else: tools like Coinglass’s liquidation heatmap map where leveraged positions probably cluster — important caveat: it’s a model built from open interest and typical leverage tiers, not a leaked order book. That prices so often run straight into those zones is what the scene calls “stop hunting”; the intent of specific actors is practically never provable — the mechanism underneath is, see Osler. What the extreme looks like, October 10, 2025 showed: by tracker data, over 19 billion dollars in positions were liquidated within 24 hours; Amberdata’s minute-level data reveals the cascade’s anatomy — 3.21 billion in a single minute, 93.5 percent of it forced selling. No chart pattern “predicted” that. But the clusters the cascade ran through were visible on every heatmap beforehand. That is the difference between fortune-telling and mechanics.

The flip side holds just as firmly: in crypto, edges decay faster. The most thorough crypto TA study (Hudson and Urquhart, 2021 — nearly 15,000 rules, with data-snooping controls) found broad profitability in the market’s early phase — and for Bitcoin it was gone in the out-of-sample period. The same pattern as Sullivan, Timmermann and White in 1999. The flock learns.

What’s left without the magic

If TA is coordination at its core — what does that mean in practice? Two things, both less spectacular than any YouTube thumbnail.

First: levels are places, not oracles. A support doesn’t “hold” because the line is sacred; it holds if demand actually waits there — and breaks if it doesn’t. Reading levels as a probability map of other people’s orders rather than as prophecy is using the tool the way the evidence supports.

Second — and this is the part that holds even if you believe in nothing: the framework. The measurable benefit of rules like stops doesn’t hang on chart magic: Kaminski and Lo showed that stop-loss rules cost you returns in random-walk markets and can add value in momentum regimes — it depends on the market, not the line. And in Detzel et al.’s Bitcoin study, simple average-based rules measurably reduced one thing above all: the depth of drawdowns. An invalidation point, a position size, a defined “I was wrong” — that is the part of TA that enforces discipline where feelings would otherwise trade. No prophecy required.

That is my honest summary of the “magic”: technical analysis works where it maps coordination, and it helps where it enforces discipline. It fails reliably the moment you use it as fortune-telling — and it gets most expensive with leverage, because then the cascades join the game. Whether and what you do with it remains entirely your call; this post has hopefully just sorted the why. The vocabulary lives in the index — and the core rules of the sober craft exist in the Crypto Collection as the Trading Cheat Sheet, a tool rather than a promise, like everything here.

Sources and limits

The research base is documented through primary sources (Journal of Finance, Review of Financial Studies, JFQA, Journal of Financial Markets, BIS; freely accessible versions linked where they exist). Important limits: the crypto studies mostly cover early, notoriously inefficient market phases — their results are not statements about today. The coordination mechanics are most rigorously documented for FX and equity markets (Osler; order clustering); their transfer to crypto is well-grounded interpretation, marked as such. “Stop hunting” as the deliberate practice of specific actors is not legally established and is explicitly treated here as scene description. And the most important limit last: nothing in this post is a trading strategy — it explains why things happen, not what you should do.

Questions, or spotted a mistake? Write to me.