Finance
The Random Walk: Why Stock Prices Are So Hard to Predict
The Random Walk is the idea that successive changes in a stock's price are statistically independent, so the best forecast of tomorrow's price is simply today's price. If markets efficiently absorb all known information, then only genuinely new information — news, by definition unpredictable — moves prices. Price changes therefore behave like a sequence of coin flips: memoryless, unpredictable in direction, and spreading out over time like diffusing particles. The result is deeply counterintuitive: the jagged charts that look full of patterns are, to a first approximation, patternless. This is why professional stock-pickers so often fail to beat a simple index fund.- First describedLouis Bachelier, 1900 (Théorie de la spéculation)
- Popularized byBurton Malkiel, A Random Walk Down Wall Street (1973)
- Core relationE[Pₜ₊₁ | info today] = Pₜ (a martingale)
- Key conditionIndependent, identically-distributed increments
- Spread grows as√t (uncertainty ∝ square root of horizon)
- Famous testCowles 1933: forecasters no better than chance
Interactive visualization
Press play, or step through manually. The visualization is yours to drive — try it before reading on.
Watch the 60-second explainer
A condensed visual walkthrough — narrated, captioned, under a minute.
The mechanism: why unpredictability is a symptom of efficiency
The random walk is not a claim that markets are chaotic or irrational — it is the opposite. It says prices are unpredictable because they are well-informed. Suppose a stock will pay off based on some future outcome, and everyone can see the same news. If it were widely known today that the price would rise 5% next Tuesday, buyers would compete for the stock now, bidding it up until the expected extra gain vanished. Predictable price moves are self-destroying: the act of predicting them removes them.
What's left is the part nobody can foresee — the arrival of genuinely new information. Since news is by definition unpredictable (otherwise it wouldn't be news), the price changes it drives are unpredictable too. Formally, the price becomes a martingale: E[Pₜ₊₁ | all information today] = Pₜ. The expected change is zero regardless of the entire past history. This is the precise, testable core of the idea — and it follows directly from the Efficient Market Hypothesis.
Note the subtlety: a martingale allows the size of moves to depend on the past (volatility can cluster), and it allows a small upward drift for risk compensation. What it forbids is a reliable rule for guessing direction.
The model, in numbers: drift, diffusion, and the √t rule
Write the change in log-price over one step as Δp = μ + σ·ε, where μ is a small drift (the average return you earn for bearing risk), σ is the volatility per step, and ε is a random draw with mean 0 and standard deviation 1 that is independent from step to step. Over t independent steps the drifts add up linearly but the random shocks add up only as a square root, because independent noise partly cancels:
Expected total move = μ·t | Uncertainty (std dev) = σ·√t
This √t is the signature of a random walk and explains why long-horizon forecasts are so wide. Take the S&P 500: roughly μ ≈ 8%/year and σ ≈ 16%/year. Over 1 year the ±1σ band is ±16%. Over 25 years the average return compounds handsomely (drift × 25), while the annualized uncertainty shrinks to 16% ÷ √25 ≈ 3.2% — which is exactly why stocks look risky over a year but reliable over decades. Drift wins the long race; diffusion dominates the short one.
The continuous-time version, geometric Brownian motion (dP/P = μ·dt + σ·dW), is the exact model Black, Scholes and Merton used in 1973 to price options — a Nobel-winning application built directly on the random-walk skeleton.
A worked coin-flip example
Strip it to the bone. Start a price at 100. Each day flip a fair coin: heads +1, tails −1. That's a symmetric random walk with μ = 0 and step size σ = 1.
- Best forecast of tomorrow: today's price. After a run of five heads (price 105), the coin still has no memory — tomorrow's expected value is 105, not "due for a fall." Believing otherwise is the gambler's fallacy.
- Spread after 100 days: standard deviation = √100 = 10. So you'd expect to be within roughly ±10 of 100 — a path landing anywhere from ~90 to ~110 is unremarkable.
- Spread after 400 days: √400 = 20. Quadruple the time, only double the spread. That's the √t law made concrete.
Now run the walk with several different random seeds from the same start: the paths fan out into wildly different futures even though the rule is identical. None of them is "wrong." This fanning cone — narrow now, widening with √t — is the honest picture of what forecasting a stock actually looks like.
The evidence — and where the walk stumbles
The empirical record is strong at short horizons. Alfred Cowles (1933) tracked thousands of stock-market forecasts from professional services and found they did no better than chance. Maurice Kendall (1953) examined 22 price series expecting cycles and found "wandering" data — his shock helped launch the modern theory. Decade after decade, the SPIVA scorecards show that after fees roughly 80–90% of active US equity funds trail their index over 15-year windows — the practical verdict of a near-random walk plus costs.
But the walk is an approximation, not a law of nature, and it frays at the edges:
- Fat tails and volatility clustering. Real returns aren't the tidy bell curve the simple model assumes. Crashes like Black Monday, 19 Oct 1987 (the Dow fell 22.6% in a day) are astronomically unlikely under a Gaussian random walk — a many-sigma event that should almost never happen. Big moves cluster in time (ARCH/GARCH captures this).
- Documented predictability. Momentum (Jegadeesh–Titman, 1993), long-run reversal, and the value effect show returns aren't perfectly memoryless. Lo and MacKinlay's 1988 variance-ratio test formally rejected the strict random walk for weekly US indices.
The honest summary: prices are close to a random walk, close enough that beating them net of costs is brutally hard, but not exactly one.
The misconception people can't shake
The single most common error is reading the past into the future — seeing a chart, spotting a "head-and-shoulders" or a "trend," and inferring what comes next. Random walks generate rich-looking patterns by pure chance. In a famous demonstration, Burton Malkiel had students chart a series of coin-flip prices; a chartist he showed it to declared it a screaming buy — it was noise. Our pattern-hunting brains manufacture signal from randomness.
A second, subtler confusion: "random walk" does not mean prices go nowhere or that investing is a coin flip with zero expected gain. There is a positive drift μ — the equity risk premium, historically ~4–6% per year above cash — that compensates you for holding risk. The claim is narrow: you cannot reliably predict the timing and direction of the deviations around that drift. You get paid to bear uncertainty, not to outsmart it.
Third: independence of returns does not require constant volatility. A martingale can have wild, clustering swings and still be unforecastable in direction. Predicting how much a price will move is very different from predicting which way.
Why it matters: index funds, options, and the limits of forecasting
If direction is unpredictable, the rational response is not to try. That logic — random walk plus the drag of trading costs and fees — is the intellectual engine behind passive index investing. John Bogle launched the first retail index fund at Vanguard in 1976, mocked at the time as "Bogle's Folly"; passive vehicles now hold trillions and often outpace the majority of active managers precisely because they refuse to pay for forecasts that don't beat noise.
The same skeleton underpins derivatives pricing: because you can't predict the path, you price an option off the distribution of paths (the widening cone), which is exactly what Black–Scholes does with σ and √t as its main levers. It shows up across markets — foreign exchange rates are famously walk-like at short horizons (Meese–Rogoff, 1983, found no model beats a random walk for FX forecasting), and short-dated Treasury moves are nearly unforecastable.
The deeper lesson generalizes beyond finance: in any competitive arena where participants act on information, the exploitable patterns get arbitraged away, and what remains looks like a random walk. Unpredictability is the fingerprint of a market that is working. See Arbitrage for the force that enforces it.
| Feature | Random Walk (Martingale) | Mean-Reverting Process |
|---|---|---|
| Best forecast of Pₜ₊₁ | Today's price, Pₜ | A weighted pull back toward a long-run average |
| Are past returns useful? | No — increments are independent | Yes — a low price predicts a future rise |
| Spread of outcomes over time | Grows without bound (∝ √t) | Bounded — variance settles to a fixed level |
| Profitable trading rule exists? | No systematic one (before costs) | Yes — buy low, sell high around the mean |
| Real-world fit | Short-horizon stock returns, FX | Interest-rate spreads, some commodities, valuation ratios over years |
Frequently asked questions
If stock prices are a random walk, why does the market go up over time?
Because a random walk can have a positive drift. The technical name is a 'random walk with drift': each step is unpredictable in direction, but the average step is slightly positive — the equity risk premium, historically around 4–6% per year above cash, paid to you for bearing risk. Randomness describes the wiggles around the upward trend, not the trend itself. Over decades the drift dominates; over days the random diffusion dominates.
Doesn't the existence of billionaire investors like Warren Buffett disprove the random walk?
Not by itself. With millions of investors, some will beat the market for long stretches by pure luck — a random walk guarantees big winners and losers even if skill were zero. Distinguishing genuine skill from luck statistically requires an enormous, persistent track record, which very few achieve. That said, the persistence of a handful of managers, plus documented anomalies like momentum and value, does suggest markets are close to, but not exactly, a perfect random walk.
What is the difference between a random walk and a martingale?
A random walk usually assumes increments are independent and identically distributed — same distribution every step, no memory at all. A martingale is weaker: it only requires the expected next change to be zero given all past information, while allowing the size of moves (volatility) to depend on history. Modern finance leans on the martingale version, because real returns show volatility clustering yet are still hard to predict in direction. Every IID random walk is a martingale, but not vice versa.
Why does uncertainty grow with the square root of time, not linearly?
Because independent random shocks partly cancel. If each step has standard deviation σ and steps are independent, variances add: after t steps the total variance is σ²·t, so the standard deviation is σ·√t. Drift, by contrast, accumulates linearly (μ·t). The √t law means quadrupling your horizon only doubles the spread — which is why stocks look terrifying over one year but far tamer, annualized, over twenty-five.
If everyone believes in the random walk and buys index funds, would markets stop being efficient?
This is the Grossman–Stiglitz paradox. If no one does research because prices are already efficient, prices would stop reflecting information and become predictable — creating profit opportunities that lure active traders back in. So the market sits at an equilibrium where just enough people trade on information to keep prices nearly (but not perfectly) unpredictable, and those traders earn just enough to cover their costs. Perfect efficiency is self-undermining; near-efficiency is the stable state.
How do technical analysts and chartists survive if patterns are just noise?
Under the strict random walk, chart patterns have no predictive power — Malkiel's coin-flip charts fooled experienced chartists into 'buy' calls. Rigorous tests find most technical rules don't beat buy-and-hold after transaction costs. Chartists persist partly because randomness produces convincing-looking patterns, partly because of survivorship and confirmation bias, and partly because markets aren't a perfect random walk, leaving thin, transient edges that are hard to exploit reliably after costs.