- Across 19,290 S&P 500 sessions, yesterday's return tells you nothing about tomorrow: the rules "everyone uses" to call the direction are right between 47% and 52.5% of the time — and after even minimal costs, every one of them returns less than simply buying and holding.
- A century of research — from Cowles (1933) to the 2013 Nobel Prize — says the same thing. And the professionals can't do it either: 88.29% of active U.S. funds underperformed the index over the last 15 years.
- Missing the 20 best sessions out of 19,290 cuts the final result from ×463 to ×115. And all 20 came in the middle of a selloff: a drawdown of more than 10% and the index below its 200-day moving average.
- Some things, though, are predictable: volatility clusters, time loads the coin in favor of those who stay invested (53.5% of days positive, 93.1% of decades positive), and regimes persist.
- It's the difference between predicting the market (impossible) and riding it with systematic rules: the conclusion isn't to give up — it's to change jobs.
Every morning thousands of analysts, newspapers, YouTube channels and newsletters churn out forecasts: where the market will go tomorrow, next month, by year-end. An entire industry makes a living selling the answer to a single question: can you predict market moves? The funny thing is that this question already has an answer. It has had one for almost a century, it is among the best-documented answers in all of economics, Nobel Prizes have been awarded for proving it — and it's no. Not "no, unless you have the right indicator." No.
As always on this blog, though, we don't ask you to take our word for it: first we briefly retrace what the research has found, then we redo the math from scratch on 76 years of S&P 500 data, with tests anyone can replicate. And at the end we get to the point almost everyone skips: if the market can't be predicted, why do systematic approaches work? The answer to that second question is worth more than the first.
A century of answers: what the research says
In 1933 Alfred Cowles III — an heir who had ruined his health searching for the right forecaster to follow — did the most sensible thing possible: he put them all to the test. His study, with the telling title Can Stock Market Forecasters Forecast?, analyzed thousands of recommendations: 16 professional financial services had on average returned 1.43% a year less than the market; 24 financial publications did 4% a year worse than pure chance. His conclusion, in his own words: the best records were "little, if any, better than what might be expected to result from pure chance".
No, and that's the striking part. S&P's SPIVA reports measure every year how many mutual funds — run by professionals paid to beat the market — actually do it: as of mid-2025, over 15 years 88.29% of U.S. large-cap equity funds underperformed the S&P 500. And CXO Advisory, which tracked 6,582 public forecasts from 68 "gurus," measured an average accuracy of 47.4% — worse than a coin flip.
On the retail side it's even worse: Barber and Odean's classic study of 66,000 accounts shows that those who trade the most systematically earn the least, and research on Taiwanese day traders (fifteen years of data, the entire market) finds that less than 1% earn predictable profits net of costs.
The theoretical explanation came in the 1960s from two future Nobel laureates. Paul Samuelson showed that properly anticipated prices fluctuate randomly: if tomorrow's rally were predictable today, someone would buy today, moving the price today — and the predictability would destroy itself the instant it was discovered. Eugene Fama formalized the idea in the efficient market hypothesis: prices already reflect available information, so whatever moves them tomorrow is, by definition, what we don't know today. The news. And the news doesn't keep a calendar.
In 2013 the Nobel committee closed the circle, jointly honoring Fama and Robert Shiller — two intellectual rivals — with a citation that reads as if it were written for this article: "there is no way to predict whether the price of stocks and bonds will go up or down over the next few days or weeks", but "it is quite possible to foresee the broad course of these prices over longer periods". The short term is unpredictable; long-term regularities are real. Keep this two-part statement in mind: it's the key to everything that follows.
Let's run the numbers: 76 years of the S&P 500 under the microscope
Test 1 — Does yesterday tell you anything about tomorrow?
If the market were predictable from its own past, daily returns would be correlated with each other: up days heralding more up days, selloffs setting up rebounds. The measure is called autocorrelation, and the chart below shows it alongside its twin computed on the size of the moves (squared returns). Look at the teal bars: all squashed inside the random band. The largest, out of twenty lags analyzed, is 0.037: the day that "explains" the most explains about one thousandth of tomorrow's variation.
Autocorrelation: direction (≈ zero) vs. magnitude (high and persistent)
Test 2 — Are winning and losing streaks "real"?
53.5% of sessions are positive (make a note of this number: we'll need it). But are the streaks — the runs of up and down days that commentators read as "momentum" or "capitulation" — any different from those of a coin? The runs test says the difference exists but is tiny: 9,044 runs observed vs. 9,537 expected from pure chance. Market streaks are only slightly longer than a coin's: a real statistical echo, but so faint that — as the next test shows — it doesn't even cover the minimal costs of exploiting it.
Test 3 — The rules "everyone uses," put to the test
Let's get concrete. Four of the most popular forecasting rules, applied with no hindsight: every evening the rule sets tomorrow's position using only data available that evening. Costs: just 2 basis points per position change — a generous, almost lab-grade assumption.
| Rule | Correct calls | Result (net of costs) | Buy & Hold same period |
|---|---|---|---|
| Tomorrow like yesterday (follow the sign) | 52.5% | ×245 | ×454 |
| 5-day momentum | 51.5% | ×20 | ×453 |
| Above/below the 20-day average | 51.3% | ×20 | ×455 |
| Contrarian RSI(14) (buy oversold) | 47.1% | ×0.84 (loss) | ×4.3 |
Test 4 — Could you spot the real market?
The most instructive test for anyone who believes they can "read charts." Below are four series of 2,000 sessions, all starting at 100. One is the actual S&P 500 over the last eight years. The other three are pure chance: numbers randomly drawn by a computer with the same mean and volatility as the market. Before reading on, pick one: which is the real one?
Series A
Series B
Series C
Series D
The answer is Series B — the one with the vertical crash a fifth of the way in (Covid, March 2020). But look at the others: C draws a textbook "double top" before plunging, A respects its "support levels," D puts together a nice uptrend. They're coin flips. Every chart pattern you think you see in the market, pure chance draws just as well — which is why your eyes are not a forecasting tool.
Test 5 — The world isn't "normal": fat tails
There's an even deeper reason why point forecasting is an illusion: the market doesn't follow the "tame" statistics of the bell curve. Let's count the extreme sessions we should have seen in 76 years if returns were Gaussian, and the ones we actually saw:
Extreme sessions: expected in a Gaussian world vs. observed (log scale)
October 19, 1987: the "impossible" event
The extreme case is the 1987 crash: −20.5% in a single day, 23 standard deviations. In a Gaussian world an event like that shouldn't have happened even once since the beginning of the universe. It happened on a Monday. Fat tails aren't a curiosity: they're the reason anyone trading on "predictions" sooner or later runs into the day they didn't see coming — and that day decides their career. Any serious method has to be built to survive the unpredictable, not to pretend to anticipate it.
Test 6 — The cost of getting the timing wrong
"Fine, not day by day — but I just want to dodge the crashes and get back in afterward." It's the most reasonable objection, and the numbers demolish it in the most elegant way. Since 1950, buy & hold has multiplied capital 463 times (8.35% compounded annually). Now let's remove, out of the 19,290 sessions, only the best ones:
76 years of the S&P 500: buy & hold vs. missing the 20 best sessions
Of course, the reverse is true too: avoiding the 20 worst sessions would have returned ×2,588. But here's the figure almost nobody publishes: anyone who managed to avoid both the 20 best and the 20 worst — in other words, anyone who simply "sat out the turbulent periods" — would have made ×642: almost the same result as doing nothing. The payoff isn't in avoiding turbulence: it's in telling, inside the turbulence, the good day from the bad one. That is, in predicting the exact day.
The plot twist: what you can know
If the article ended here, the conclusion would be depressing: it's all chance, so you might as well be a fatalist. But remember the Nobel committee's two-part statement? Short-term direction is unpredictable, and yet there are regularities over longer horizons. Our own data confirm it, in three precise ways.
Regularity #1 — Volatility clusters
Go back to the first chart in the article. The teal bars (direction): all at zero. The gold bars (the size of the moves): all high, positive, persistent — with overwhelming statistical significance. Translation: knowing the market went up today tells you nothing about tomorrow; knowing the market was turbulent today tells you a lot about how turbulent it will be tomorrow. Turbulent days come in clusters — our tables of extreme events are all bunched in 1987, 2008 and 2020 — and so do calm stretches. Direction is unpredictable; the "climate" isn't. Anyone who trades options knows what this asymmetry is worth: you can estimate the price of uncertainty, not its direction.
Regularity #2 — Time loads the coin in your favor
Percentage of positive periods, by time horizon
It isn't magic: it's the equity risk premium — the compensation the market pays, over time, to those willing to sit through its swings. On any single day the coin is almost fair; over a decade it's overwhelmingly loaded. The only "forecast" that has always paid isn't about tomorrow's direction: it's about the direction of patience.
Regularity #3 — Regimes persist
One last piece. Let's split the 19,290 sessions using the simplest possible criterion: was the index above or below its 200-day average yesterday? No forecasting — just a snapshot of where we are. And yet the two groups describe two different markets:
Two markets: annualized return and volatility, above and below the SMA200
Today's regime doesn't predict tomorrow — but it dramatically changes tomorrow's odds and turbulence. Knowing what climate you're in isn't knowing what the weather will be: it's knowing whether to take an umbrella.
Not predicting, but riding: what changes in practice
Let's recap the apparent paradox: short-term market direction can't be predicted, yet volatility, the risk premium and regimes have been solid regularities for seventy-six years. The conclusion isn't "nothing can be done": it's that the right job isn't forecaster — it's insurer. An insurer doesn't know which house will burn down; nobody does. Yet it makes money systematically, because it doesn't bet on the single event: it collects a premium calibrated to the probabilities, diversifies, and defines in advance the maximum loss it can withstand. Replace "which house will burn down" with "what the market will do tomorrow" and you have the exact definition of the systematic approach.
✕ What doesn't work: predicting
- Guessing the direction of tomorrow, the week, the quarter
- Getting out "to be safe" during selloffs and back in when the sun comes out: the best days live inside the storms
- Reading chart patterns that pure chance draws just as well
- Trading without a defined maximum loss, exposed to the 23-sigma day
- Making decisions under stress, session by session, headline by headline
✓ What works: riding the market
- Staying exposed to the risk premium, which pays in 74% of years and 93% of decades
- Regime rules (like the 200-day average) that reduce exposure to turbulence
- Exploiting volatility clustering: the price of uncertainty can be estimated, the direction can't
- Defined risk on every position: surviving the unpredictable is the prerequisite for everything
- Mechanical protocols written in advance, always executed, tested on decades of data
The verdict
Predicting market moves isn't possible: not for you, not for the professionals (88% of funds trailing the index over 15 years), not for the gurus (47% accuracy). Ninety years of research, two generations of Nobel laureates and the 19,290 sessions sifted through in this article all say so. But the same analysis that closes the door on prediction opens the one next to it: you don't predict the market, you ride it. You collect the premium it pays to those who stay, you respect the climate it signals, you estimate the price of uncertainty, and you build every position to survive the day that nobody — nobody — will see coming. Whoever stops asking "what will the market do tomorrow?" and starts asking "how am I positioned for whatever it does?" has already made the most important shift of their investing life.
This is exactly the foundation all of Kriterion Quant's work is built on: the model portfolio's systems don't predict — they follow tested rules, with defined risk and a public track record. If you want to see how these principles turn into numbers, the place to start is the comparison between simple signals and buy & hold or the study on systematic put selling.
How we ran the numbers
Methodology. Daily S&P 500 series (price index, EODHD data), Jan 3, 1950 – Sep 4, 2026, 19,290 sessions; log returns on adjusted closes. Autocorrelation with 95% confidence band and Ljung-Box test; Wald-Wolfowitz runs test on signs; forecasting rules evaluated strictly out-of-sample (position set the evening before, costs of 2 bp per position change); simulations with the same mean and standard deviation as the real data, fixed and reproducible seed; count of extreme events against the Gaussian expectation; timing scenarios built by excluding extreme sessions; regimes defined on the prior day's SMA200 (no look-ahead). In fairness: some deviations from pure randomness are statistically detectable — Ljung-Box picks them up, and so does the Lo-MacKinlay (1988) variance ratio: markets are not a perfect random walk. The thesis isn't that structure doesn't exist: it's that it isn't structure a forecaster can use — too small to survive costs, and concentrated in the size of moves, not their direction. Every number can be recomputed: the test notebook is available on request.
Main sources (verified): Cowles (1933), Econometrica · Samuelson (1965), Industrial Management Review · Fama (1970), Journal of Finance · Nobel 2013, nobelprize.org · Barber & Odean (2000), Journal of Finance · Barber, Lee, Liu, Odean (2014), Journal of Financial Markets · SPIVA U.S. Mid-Year 2025 · CXO Advisory, Guru Grades · Lo & MacKinlay (1988), Review of Financial Studies.
This article is also available in Italian.