The best reading on technical analysis is the record of it being tested

If you want to learn to read charts, the usual advice is a stack of technical analysis books. This is a different list, and the argument for it is simple: technical analysis has been tested seriously by economists for more than thirty years, and the record of that testing is more useful than any of the manuals.

Not because the manuals are worthless — there are four below and each earns its place. But a manual tells you what a pattern is called. The research tells you what happened when someone checked, which is the question you actually had.

Three tiers, in the order worth reading them.

Tier 1 — the review that tells you where the whole literature landed#

Park, C.-H., & Irwin, S. H. (2004/2007), The Profitability of Technical Analysis: A Review.

Start here, because it saves you a year. It surveys the empirical literature and sorts it into "early" and "modern" studies by the quality of the testing procedure.

Read the tally and then read the sentence that follows it, because the second one is the point: most of those studies are subject to data snooping, ex post selection of trading rules, and difficulty estimating risk and transaction costs. A count of positive results is not a verdict; it is a description of a literature with a known set of weaknesses.

⚠️ A note on which version to cite. The freely readable 2004 working paper counts 92 modern studies as 58 / 24 / 10. The 2007 journal version counts 95 as 56 / 20 / 19. We quote the working paper because it is the one you can open.

Tier 2 — the five papers that are the actual story#

If you read nothing else, read this sequence. It is a complete arc: a positive finding, a rigorous check that confirmed it, and then an out-of-sample result that receives much less attention.

1. Brock, Lakonishok & LeBaron (1992). Two of the simplest and most popular rules — moving averages and trading-range breakouts — tested on the Dow from 1897 to 1986, with bootstrap methods extending the standard statistics. The results supported the rules, and did not fit four null models including the random walk and GARCH variants. It became one of the landmark empirical papers in the modern academic study of technical trading rules.

2. Sullivan, Timmermann & White (1999). The obvious objection to any such finding is that if you try enough rules, some will look good by luck. They applied White's Bootstrap Reality Check to about 8,000 rules drawn from five trading systems, across 100 years of the Dow.

And Brock, Lakonishok and LeBaron survived it. Within the 1897–1986 period, the best rule from that 26-rule universe returned an annualised 9.4% with a data-snooping-adjusted p-value of zero, against 4.3% for buy-and-hold. The finding was not an artefact of searching.

3. The same paper's out-of-sample section, which is the part that matters. They kept ten more years — 1987 to 1996 — that Brock, Lakonishok and LeBaron had not seen. The best rule selected on the earlier data returned 2.8% a year with a p-value of 0.32. In the review's summary, it "did not continue to generate valuable economic signals in the subsequent period."

A later expansion to 17,298 rules put the best recent rule at 17.3% against 13.6% for buy-and-hold — with a data-snooping-adjusted p-value of 0.98.

4. The cost line. Separately, work summarised in the same review found that for the 1976–1991 subsample the break-even one-way transaction cost of these rules had fallen to 0.22%, against estimated actual costs of 0.24%–0.26%. The edge had not vanished so much as become smaller than the toll for using it.

5. Lo, Mamaysky & Wang (2000), Foundations of Technical Analysis. This one points the other way and belongs on the list for exactly that reason. They built an automatic pattern recogniser using nonparametric kernel regression, applied it to US stocks from 1962 to 1996, and compared the unconditional distribution of returns with the distribution conditioned on classical patterns — head-and-shoulders, double bottoms and the rest. Their conclusion is that several technical indicators do provide incremental information.

⚠️ That is a positive result from a serious paper, on daily bars, about information content. It is not the same claim as "these rules are profitable after costs", and it does not contradict the sequence above. Different horizon, different question. A reading list that left it out would be selecting its evidence, which is the failure mode this whole article is about.

And a dissent worth knowing exists: Ready (2002) argued that Brock, Lakonishok and LeBaron's results were spurious. Park and Irwin note the disagreement rather than resolving it. So should you.

The four things that arc separates#

This is the reason the sequence is worth more than any single result in it. In-sample pattern, statistical significance, economic usefulness, and out-of-sample validity are four different things, and almost every argument about charts confuses at least two of them.

• In-sample pattern — the rule looks good on the data you have.
• Statistical significance — the pattern is unlikely to be chance given how much you searched. Brock, Lakonishok and LeBaron cleared this; that was the contribution of the 1999 paper.
• Economic usefulness — what survives after costs. The 0.22% against 0.24–0.26% line.
• Out-of-sample validity — it still works on data nobody had seen. The 2.8% at p = 0.32.

A finding can pass the first three and still fail the fourth, and the literature above contains important examples of exactly that.

Tier 3 — books worth reading anyway#

Not because they establish that anything predicts. Because they give you language, categories and a sense of what other people are looking at when they look at a chart — and none of that is available from a paper.

Each entry is what to take, and what not to conclude.

Murphy, Technical Analysis of the Financial Markets (New York Institute of Finance, 1999). Take: the shared vocabulary and classification system of the whole field, in one place. If you want to read anything else — including the papers above — you need this taxonomy first. Do not conclude: that anything in it has been tested in the form it is presented.

Nison, Japanese Candlestick Charting Techniques, 2nd ed. (Prentice Hall Press, 2001). Take: the candlestick vocabulary and, more interestingly, its history. Knowing where these terms came from is part of understanding why so many of them are defined so loosely. Do not conclude: that a named candle formation carries a tested edge. We tried to measure one of the most famous of them and could not even get four published definitions to agree on which bars counted — 39.6% overlap.

Edwards & Magee, Technical Analysis of Stock Trends (Stock Trend Service, 1948; 11th ed., CRC Press, 2018). Take: the original chart-pattern taxonomy, the source most later books are downstream of. Do not conclude: prediction. But note the connection — the patterns Lo, Mamaysky and Wang formalised in 2000 are largely this book's patterns. The book supplied the vocabulary that made the test possible fifty years later. That is a real contribution and it is not the same one the book claims for itself.

Steenbarger, Enhancing Trader Performance (Wiley, 2006). Take: this is the only book of the four about how anyone gets better at anything, and it is built explicitly on the expertise research — deliberate practice, immediate feedback, structured repetition rather than accumulated screen time. It is the bridge between this reading list and the learning-science literature that is more directly relevant to whether reading the other three turns into practice. Do not conclude: that it will tell you what to trade. It is not trying to.

A sidebar, not a fifth entry#

Lefèvre, Reminiscences of a Stock Operator (1923) — originally twelve articles in The Saturday Evening Post between June 1922 and May 1923, a roman à clef built on Jesse Livermore.

It sits outside the tiers on purpose. Its value is narrative and psychology, and it is evidence of nothing. Read it the way you would read any first-person account written a century ago: for what it was like, not for what is true. The behavioural pattern it describes best has since been measured properly — Odean (1998) found across 10,000 brokerage accounts that investors realise winners far more readily than losers, which is the same effect from the other side.

Where our own work sits in this#

We are not claiming the last thirty years of researchers were wrong. We keep arriving at the same wall they did, at a much shorter horizon, with far less at stake.

Our own naive contrarian rule is a coin over 366,606 unfiltered decision points. Five confirmation rules — volume, trend, VWAP and two forms of breakout — all cleared p below 0.0001 on 1.13 million observations, and all were worth about half a percentage point, which is the statistical-significance-versus-economic-usefulness distinction arriving unannounced. And a headline figure we published ourselves turned out to be produced by a de-duplication step in our own pipeline rather than by anything in the market — our own version of the in-sample-versus-out-of-sample problem, wearing different clothes.

That is why the reading list is annotated rather than ranked. We have publicly mistaken an in-bank result for a market result, treated statistical detectability as more meaningful than effect size, and misread a sampling artefact as a market pattern — three of those four distinctions, in public, in a fortnight. The literature above is a record of much better resourced people meeting the same ones.

What this list does not do#

It does not tell you what to trade, and none of the papers do either. It does not settle whether technical analysis works — the honest summary of thirty years is that some rules produced results that survived rigorous in-sample tests and then did not persist, and that reasonable researchers still disagree about parts of it.

And it will not, by itself, make you better at reading a chart. Steenbarger's book is on the list because it is the one that says why: reading is not repetition, and expertise comes from repeated attempts with fast feedback. That is a practice problem, not a bibliography problem, and it is the one thing on this page a reading list cannot solve.

SwipeTA is a training game and a simulation: no real money, no broker, and it does not provide investment advice. The parameters behind our own measurements are on the methodology page.