Technical Analysis for Stocks: Worse Than Random on 30 Names

I ran the 50/200 moving average rule on 30 large American stocks, one at a time, for almost 26 years. The median name turned $10,000 into $31,030. Buying the same stock on day one and ignoring the chart turned it into $78,547. That gap is the short version of why technical analysis for stocks is a weaker tool than most people assume: on the median holding, the rule cost about 60% of the result.

Then I ran the control that changed my mind about what was going on. Instead of following the signal, I held each stock on randomly chosen days, matching the number of days the rule was invested, 500 draws per name. The rule landed at the 34th percentile of that random distribution for the median stock, and beat random on only 10 of the 30 names. It was not merely unprofitable. On most of these companies it picked worse days to be invested than chance did.

My first instinct was that this is a single-company problem: one stock is mostly its own noise, so average enough of them together and the signal should start working. I tested that too, and it is wrong. Equal-weight baskets built from the same 30 names, at every size from 3 to 30, left the rule between 3.2 and 4.1 points a year behind simply holding the basket, and still sitting around the 22nd to 36th percentile of random. Diversification did not rescue the chart.

I want to be clear about what I am and am not arguing, because “technical analysis does not work” is not my position. My position has two halves. A chart is a short-horizon instrument. Inside one session it describes something real and immediate, and that is where I think it earns its keep: daytrading, where you live inside a continuous auction with a hard stop and you go home flat. Over years, the information sits somewhere else entirely — in the business, and in where the market’s money is actually going. Both halves matter, so the rest of this piece tests the first half and then points at what replaced it in my own work.

The test I froze before looking at anything

Every number here comes from daily bars pulled from Yahoo Finance, 3 January 2000 to 18 September 2026. After the 200-day warmup the averages need, that leaves 6,518 trading days and 25.9 years per name. I wrote the rules and picked the names before looking at a single result, because the easiest way to prove anything about charts is to keep nudging a parameter until the equity curve smiles at you.

The universe is 30 large, liquid American companies that were already large in 2000: Apple, Microsoft, Johnson and Johnson, Coca-Cola, Exxon, GE, Intel, Cisco, IBM, Pfizer, Walmart, Procter and Gamble, JPMorgan, Bank of America, AT&T, Verizon, Disney, McDonald’s, Home Depot, Chevron, Merck, Amgen, Nike, Oracle, Qualcomm, Texas Instruments, Caterpillar, Boeing, 3M and Ford. Some of them multiplied many times over. Some went nowhere for a quarter of a century. I wanted both.

This article stays inside that universe. No index, no ETF, no futures, no currency pairs. An earlier draft of this piece compared the stocks against the S&P 500 and a few FX pairs, and I cut it: swapping the instrument changes the weighting, the participants and the structure of the market all at once, so a difference in the result tells you nothing about stocks. Everything below compares stocks with stocks.

Returns use the dividend and split adjusted series, so holding gets full credit for its dividends and the rules are not flattered by a total return being compared against a price return. Every rule pays 0.05% per side, which is generous for a modern broker and nowhere near enough to explain the results. Signals are computed on the close and acted on at the next close, so nothing trades on information it did not have. Gap and stop measurements use the raw open, high, low and close, because that is what a chart draws. Idle cash earns nothing, which I come back to in the caveats, and there is no tax, which flatters the rules rather than the holder.

Technical analysis for stocks lost to doing nothing on 27 of 30 names

Three rules, thirty companies, one table.

After costs, across the same 30 names. Three things are worth sitting with.

First, the size of the miss. Four points of annual growth over 26 years is not a rounding error, it is most of the outcome. The median stock held to the end produced $78,547 from $10,000, and the same stock traded on the 50/200 rule produced $31,030.

Second, the long and short row. This is the honest test of whether a chart predicts direction, because it removes the crutch of being long a rising asset. Betting both ways on the breakout lost 6.62% a year, with an average worst fall of 88.8%, and was profitable on 3 of the 30 names. If a breakout on a single company carried directional information, trading it in both directions should not do that.

Third, the drawdown column is the one honest win. Every long-only rule sat in cash for part of every crash, so every long-only rule fell less than the stock did. The 50/200 version cut the average worst fall from 65.4% to 52.2%, and it was shallower on 25 of the 30 names. That is a real product. The question is its price.

The three names where the rule beat holding are the interesting part. GE, Cisco and Intel: one industrial conglomerate that spent two decades unwinding, and two technology companies that peaked in the dot-com bubble and then ground lower for years. The rule did not outsmart those charts. It sat in cash while three businesses deteriorated, which is a slow enough process that a reader of the annual reports would have got there first. Qualcomm is the other end of the range, giving up 9.9 points a year, because it spent the 2010s in long noisy swings that whipsawed a 200-day average without ever breaking the underlying growth.

The control that settles it: random days did better than the chart

Losing to buy and hold is not damning on its own. Any rule that sits in cash for 36% of a rising market will finish behind, and that says more about the market rising than about the signal. So I ran the control that removes the excuse.

For each stock I counted how many days the 50/200 rule was invested, then held that same stock for the same number of days chosen at random, 500 times, with a fixed seed. Both versions own the same company for the same length of time. The only difference is whether the days were picked by the chart or by a random number generator. If the signal knows anything, it should win.

It did not. For the median name the rule landed at the 34th percentile of its own random distribution. It fell in the bottom half on 20 of the 30 stocks and in the bottom quarter on 12. Averaged across the universe the rule made 4.67% a year while random days with equal time invested made 5.78%.

Diversifying the same 30 names did not rescue the signal

Here is where I expected to find the escape hatch, and did not.

The standard defence of trend following is that it needs a diversified thing to work on: one company is mostly idiosyncratic noise, average enough of them and the noise cancels, leaving something a trend rule can grip. It is a reasonable story. So I built equal-weight baskets from these same 30 stocks, at sizes 1, 3, 6, 10, 20 and 30, drawing 200 random baskets at each size below 30, and ran the identical rule and the identical random control on each basket.

Read the last two columns. Diversification did nothing for the signal. If anything the percentile got slightly worse through the middle of the range, bottoming at the 22nd percentile for baskets of ten. The return shortfall narrowed a little, from 4.08 to 3.21 points, but that is mostly the arithmetic of a rule that sits in cash less often in a smoother series, not a signal that started working.

I checked the weighting too, since equal weight with daily rebalancing is not what anybody actually holds. Buying the 30 names once and letting the weights drift gave 13.08% a year held against 9.95% for the rule, at the 53rd percentile of random, which looks like the one result that clears the coin. It does not survive a single deletion: take Apple out of that drifting basket and holding gives 9.14%, the rule gives 5.63%, and the percentile collapses to the 20th. The one apparently positive reading came from owning Apple, not from reading a chart.

The chart only helped on the companies that were falling apart

The per-stock version of the random test is where the pattern actually lives. Sort the 30 names by how well the rule did against its own random control and the split is not about volatility, sector or size. It is about whether the business worked.

The ten names where the rule beat random averaged 6.5% a year when simply held. The twenty where it lost to random averaged 9.9%. That is the whole story of what a trend rule is on a single stock: a slow detector of businesses in decline. Where a company was deteriorating over years, getting out and staying out added value, and the chart was a lagging way of noticing something the filings said earlier. Where a company compounded steadily, every exit was a mistake, and the rule spent 26 years selling McDonald’s pullbacks.

That is a genuinely useful thing to know, and it is not what technical analysis is usually sold as. Nobody puts “reliably identifies your worst holdings, several months late” on the box.

Ten days out of 6,500 decided the whole thing

There is a second reason being out of the market is expensive on a single company, and it has nothing to do with the signal.

For the median stock in this basket, the 10 best days out of roughly 6,500 delivered 61.9% of the entire 26-year result. On 19 of the 30 names, ten days carried more than half of everything. And most of the money arrived while the exchange was shut: splitting the total-return series into the overnight move and the session move, the overnight leg accounted for a median 68.8% of the whole result, and more than half of it on 17 of the 30 names.

This is the single-stock version of the arithmetic I found on the index in the study on time in the market versus timing the market, where missing the 10 best days since 1950 cut $4.59 million to $1.99 million. On one company it bites harder, because there is no diversification to spread out the arrival of good news.

What gaps actually break, and what they do not

Now the mechanical problem, and a result that cuts against the explanation I would have reached for.

These stocks move a lot while the exchange is closed. The New York Stock Exchange is open six and a half hours a day; earnings come out in the other seventeen and a half. For the median name in this universe, 36.7% of the average day’s movement had already happened before the opening bell, and the median stock opened more than 2% away from its previous close on 4.78% of sessions. The spread across the 30 names is wide: Coca-Cola averaged a 0.40% overnight move and jumped more than 2% on 1.82% of days, while Qualcomm averaged 0.92% and did it on 10.48% of days, one session in ten.

That is what breaks a stop loss. I placed a long entry at every fifth close, put a stop 7% below it, allowed 60 sessions, and recorded 19,478 stop exits across the 30 names.

So roughly one stop in seven was not honoured at its price, and in the bad tail the stock opened a further 5% to 20% below the level you chose. If your plan depends on knowing your loss in advance, that plan does not survive contact with a single company’s earnings calendar. This is the practical content of “the pattern you drew is not the trade you get.”

Here is the part that surprised me. I assumed gaps were also why the signal fails, so I checked it across the 30 names: if gaps break trend following, the gappiest stocks should be where the rule does worst. The rank correlation between a name’s average overnight move and the rule’s percentile against random is +0.46, the wrong sign for that story. Split the universe at the median gap size and the gappier half averaged the 45th percentile while the calmer half averaged the 28th.

I do not read that as “gaps help.” The gappiest names here are largely the troubled ones, Qualcomm, Cisco, Intel, Bank of America, so this correlation is most likely picking up the decline story from the previous section rather than anything about gaps. What it does rule out is the tidy version of the argument. Gaps explain why risk control fails on a single stock. They do not explain why the signal carries no information, and I should not claim they do.

The longer you intend to hold, the more wrong the signal gets

If the objection so far is “fine, but a chart is a short-term tool,” I agree, and the data agrees more than I expected. I measured how well a simple trend reading, how far the price sits above or below its 200-day average, ranked future returns at five horizons. The number is a rank correlation, where zero means no relationship.

Every horizon is negative on average, and the reading gets more negative as the horizon lengthens, which is the opposite of what you want if you are using a chart to decide what to own for years.

Be careful how much weight you put on this. The magnitudes are tiny, the windows overlap so any apparent significance is inflated, and a rank correlation of minus 0.05 is not a licence to short strength. Read it as “the chart carries no usable information about a single company, least of all over a long horizon,” not as a signal to run in reverse.

But notice which end of that table is least bad. The reading is closest to useless at one session and steadily worse the further out you go. That is the shape you would expect from a tool being used past its range, and it is the whole reason I keep charts on the screen for one job and off the screen for another.

The chart’s natural habitat is one session, not one decade

This is the half of my argument the data above does not test directly, so let me make the case plainly and then mark its limits.

Inside a single trading session, a chart is describing something that genuinely exists. The company has not changed since the opening bell. No earnings landed, no guidance moved, no product shipped. What is actually happening in those six and a half hours is an auction: inventory being moved, a large order being worked, a cluster of stops sitting under an obvious level, an opening imbalance unwinding. Price and volume are not a proxy for that process, they are that process, recorded. A daytrader reading order flow and levels is reading the only thing that is changing. Stretch the same tool to a five-year holding period and it is a rounding error next to whether the company grew its earnings.

The strongest evidence I have for this comes from the gap section above, read backwards. Two facts from these 30 stocks:

So the same measurement that makes me sceptical about charts for investing is what makes the intraday case coherent. The gap is the long-term holder’s problem and the daytrader’s non-problem. The horizon table points the same way, with the trend reading least negative at one day and most negative at one year.

What actually decided the result was which company you owned

Here is the number that reframes the whole exercise. Of the 30 stocks, $10,000 in Apple became $10,454,111. The same $10,000 in Ford became $13,502. That is a spread of 29.7 percentage points of annual growth between the best and the worst name in a basket of household-name American blue chips.

The 50/200 rule, across all 30 names, moved the annual result by between minus 9.9 and plus 2.8 points. The decision of which company to own was worth roughly three times more than the best thing the chart managed, and ten times more than what it did on average. And the chart could not help you make that decision. I checked 2003 on both names, and Ford’s was the better looking of the two.

This is why the long-horizon investor’s habit of ignoring the daily wiggles is not laziness, it is a correct reading of where the information sits. The wiggles are mostly other people’s liquidity needs. The earnings power is what compounds. Margins, cash generation, reinvestment, competitive position and the price you pay for all of it decided the 29.7-point spread in this universe, and none of those appear on a price chart. The same concentration of outcomes in a few names is why testing anything on today’s index members flatters your results.

The other thing that worked over years was sentiment, not chart reading

Fundamentals are half of my answer. The other half is sentiment, and I mean something specific by that: not surveys or magazine covers, but revealed conviction, where the market’s money is actually going, measured across many names at once.

I make this distinction because a fair objection to everything above is that I just spent 5,000 words attacking price-based rules, and the best-performing strategy on this blog is also built from prices. It is, and the difference in how it uses them is the whole point. The all-time-high momentum strategy scans the universe each month for companies that set a new all-time high in three consecutive months, each peak at least 5% above the last, then holds up to five of them for two months. Over April 2001 to December 2025, across 30 randomised universes, it averaged 11.09% a year against 8.35% for the S&P 500, and its worst drawdown was 25.08% against the index’s 47.12%.

Those are not variations on one idea, they are opposite uses of the same raw material. The rules in this article ask a single chart to forecast its own future, which is the thing my random-day control says it cannot do. The staircase filter never asks that. It asks who else is buying, with enough conviction to pay up three months running, and it answers by ranking companies against each other. That is a sentiment measurement expressed in prices, and it is closer in spirit to reading the crowd than to reading a candle.

Two honest caveats, because I am citing my own work. That study is on the index universe with randomised constituent samples, not on these 30 names, so it is a separate piece of evidence rather than a direct comparison. And a cross-sectional momentum edge of roughly 2.7 points a year is a real but modest thing that has had long stretches of underperformance, including the 2018 growth scare. It is not a machine that prints money, and the survivorship caveats apply there as much as here.

The practical upshot: for a holding I intend to keep for years, my inputs are the filings and the price I am paying, plus a cross-sectional read on where conviction is concentrating. The 50-day average of that one stock is not on the list.

Six ways this could be wrong

A study that only supports its own conclusion is a brochure. Here is where I think this one is vulnerable.

The universe survived. All 30 companies still trade. Nothing here went to zero, and bankruptcy is exactly where a trend exit earns its fee. That biases the whole comparison against the rules. You can see it in the results: the names where the rule won, GE, Cisco, Intel and Merck, are the ones that deteriorated longest, and a universe with delistings would have more of those. Rerunning this on a survivorship-free universe would make the rules look better, and it is on my list. I do not think it reverses a four-point gap, but I cannot prove that yet.

The last 16 years were a bull market. Split by decade the picture changes. Between 2000 and 2010 the 50/200 rule beat holding on 14 of the 30 stocks and added 0.93 points a year, which is what two crashes in one decade will do for an exit rule. From 2010 to 2026 it lost to holding on 30 of 30, giving up 6.18 points a year. Any rule that goes to cash looks bad in a market that mostly rises, and my full-sample verdict is an average of those two very different regimes.

Idle cash earns nothing here. The 50/200 rule sits in cash 36% of the time and I paid it 0%. At a realistic 3% that would add roughly a point a year to the rule, which narrows the gap without closing it. It does not touch the headline finding at all, because the random-day control also holds cash 36% of the time at 0%, so both sides of that comparison are treated identically.

Drawdown is a real product and I am underselling it. The rule cut the worst fall on 25 of the 30 stocks, from an average of 65.4% to 52.2%, and on the 30-stock basket from 46.0% to 33.9%. Some people should pay for that, and a few points of annual growth is a defensible price for sleeping through 2008. My objection is not that it is worthless. It is that you should know you are buying comfort, not returns, because the signal underneath is not adding information.

Three rules is not all of technical analysis. I tested a trend filter, a breakout and an oscillator. I did not test head and shoulders, Elliott waves, volume profiles or anything discretionary. I would argue these three are the most mechanical and most favourably documented family, and the drawn patterns are hard to test precisely because they get drawn after the fact. But “you did not test my pattern” is fair, and the honest answer is: not yet.

I cannot fully explain the failure. I ruled out the explanation I started with. It is not idiosyncratic noise, because averaging the names together changed nothing, and it is not gaps, because the gappiest names were where the rule did relatively best. What is left is the decline-detector pattern plus the concentration of returns in days you have to be present for. That is a description more than a mechanism, and I would rather say so than dress it up.

What I do with this

I have not deleted my charts. I have split them by horizon.

They stay on the screen for intraday work, where the session is the whole information set and I go home flat, so the gap that wrecks everything above cannot reach me. They come off the screen for anything I intend to own for years, where the chart is describing other people’s liquidity needs while the thing that compounds is happening in the accounts.

For a company I intend to own for years, I do not use a chart to decide whether to own it, and I do not use one to decide when to leave. The decision is the business and the price I am paying for its earnings. The daily bars are the noise that other people’s liquidity needs create around that, interrupted a few times a year by the only information that mattered, arriving while the market was shut.

What replaced it is the pair I described above: the filings and the valuation for the business itself, and a cross-sectional read on where conviction is concentrating for the question of what to own now. If I want drawdown control on a long-term holding, I treat it as insurance with a visible premium, roughly three to four points of annual growth in this sample, rather than as an edge. And I take seriously the one thing the data did support: when a trend rule keeps telling me to get out of a specific holding for months, that is worth reading as a question about the business, then answered with the filings rather than with the chart.

The honest summary of technical analysis for stocks is not that charts lie. It is that a chart is a short-horizon instrument, and a single company’s chart did not contain information about that company’s future at any horizon I measured, in any portfolio size I built from these names. Twenty-six years, thirty companies and three rules later, following it was worth slightly less than rolling dice. Use it where the clock is measured in minutes. For the decades, read the business and read the crowd.

FAQ: technical analysis for stocks