Does technical analysis really work? What we measured, what the research says, and what it's actually worth

The short answer: Technical analysis works partially, and mainly not in the role it's sold for. As a tool for predicting where the price is going, most of the popular rules we tested didn't beat simply buying the index, and some of them lost to it by a lot. As a tool for deciding where you're wrong, how much to risk and when not to be in the market, it has value that can be measured, and it's bigger than any single indicator.

This page doesn't try to persuade one way or the other. It presents what we measured ourselves on 184 stocks over three years, what academic research found over decades, and the reasons technical analysis looks like it works more than it actually does. All with the limitations next to the numbers, and without promises.

Why would there be anything to it at all

The assumption behind technical analysis is that price doesn't move completely randomly: trends tend to persist a bit longer than randomness would expect, and levels where the market has stopped in the past change the behavior of buyers and sellers when it returns to them. Both of these assumptions are partially supported by research. The momentum effect— that stocks which rose over the last 12 months continue on average to beat those that fell — has been documented since 1993 (Jegadeesh and Titman) and found in many markets, but it too suffers from sharp crashes. Regarding indicators and patterns, the picture is more mixed.

A 2007 literature review (Park and Irwin) found that most modern studies on technical trading rules reported positive results, but many of them weren't cleaned of multiple-testing and transaction-cost issues. When these are cleaned up (Sullivan, Timmermann and White, 1999), much of the advantage disappears. The cautious conclusion of the literature: there are weak and unstable patterns in the market, and most of what looks like an edge on first examination is noise selected after the fact.

What we measured ourselves: six methods versus the index

Instead of relying on research on American indices from the 1980s, we ran the classic rules of the popular methods on the same data: 184 American stocks and ETFs, daily candles, May 2023 through September 2026, execution at the open following the signal, 0.05% cost per side, long only, no leverage. Benchmark: buy and hold S&P 500.

• Only two methods beat the index by a small margin— and it's not certain they'd survive a different period: the turtles at 3 points, the moving-average crossover at 15. Momentum beat it by a lot, but 114 out of 188 its gains came from a single year, 2024.
• The two most popular indicators in awareness— RSI in the mean-reversion version and the stochastic — earned 14% and19% in a rising market 95%They were "right" on most of the trades, and lost to the index by a wide margin.
• The 200 moving-average filter didn't increase the return, it decreased it (74% versus 95%). What it did do: cut the maximum drawdown to11% from19%That's not prediction, that's risk management, and that's where it worked.

What else we measured, and dropped

• a trailing exit on channel break (the turtle idea) on the breakout setups of our engine: bigger runs, lower success rate, average profit per trade that didn't improve, and a worse result in an account with a limited number of positions. We stuck with a fixed target. Details inthe turtle method guide.
• Short setups: measured as losing on our side, both in the live track record and in a two-year backtest. The engine no longer shows them. Why.
• "Narrow market breadth predicts breakout failure": A claim we wrote ourselves with confidence, until we tested it on five years of data and it didn't hold up. We removed the causal phrasing and kept only the historical positioning.
• "Confidence" score for the setup: In the sample collected so far we found no correlation between the score the engine gives and the outcome. We did not calibrate the weights to fit the past, because that's the sure way to overfitting. The score remains a description of structure, not a forecast.

Why technical analysis looks like it works more than it actually does

• The chart in hindsight: On a historical chart every level "worked," because only the levels that worked get marked. In real time there are ten levels, and the price respects two.
• Impressive success rate: RSI2 was right in-62% of the trades and made less than all of them. Profit equals the win rate times the size of the win minus the loss rate times the size of the loss, and whoever sells you "80% success" doesn't tell you the other half.
• Survivorship: Whoever teaches technical analysis is someone who made money, at least for a while. Whoever lost doesn't open a channel.
• Indicators as confirmation: In a strong uptrend RSI stayed in "overbought" for weeks. Whoever sold at 70 got out at the start of the move. The indicator wasn't wrong, using it as a predictive rule was.
• One good year: Every method looks brilliant in a year that suits it. Momentum in 2024 is our example. Don't trust a table without a year-by-year breakdown.

So what does technical analysis actually give you

When you remove the prediction from the picture, three things remain that were measured as having value:

• A place to be wrong in. A stop below the structure turns "I think it'll go up" into "If it drops below 47 I was wrong, and it costs me 500." Without defining the mistake there's no risk management, without risk management there's no Position size, and without position sizing the exact same list of trades can be profitable or losing. We measure this explicitly: the size changed the sign, not the signal.
• regime. Being less exposed to the market when the index is below200 average gained in return and saved about half the decline. For someone who can't live through a drop of 19%that's the best trade in the table.
• Discipline that can be tested. A written rule can be measured, fixed, and discarded. A "feeling on the chart" can't. The value of systematic technical analysis isn't that it's more right, but that you can know how right it is.

How to test it yourself instead of just believing

Rules written before the data, realistic execution with costs, a split into development and validation periods, at least two hundred trades, and return relative to the index rather than to zero. The full protocol, including the bugs we caught ourselves, is in thebacktesting guide. And whoever promises you signals without a full public track record, including the losses, is asking for faith. More on this inAI trading signals.

How this is implemented in TechAnalyzer Pro

The system doesn't predict. It marks structure (trend, support and resistance, breakouts), proposes a setup with a stop and target when there is one, and publicly tracks what happened to every setup it generated, including the losing ones. Every change to the rules is measured before it goes live, and whatever measures as negative gets removed, even when it's popular. The full track record can be seen on theperformance page. It's not pleasant in every period, and that's exactly the point.