The Coin That Isn't Quite Fair: Cracks in the Random Walk

For decades, the reigning belief in finance was blunt and a little deflating: stock prices move completely at random. Tomorrow's move is a fresh coin-flip, unconnected to today's. Under this "random walk" view, studying charts for patterns is like studying a roulette wheel for patterns, you're fooling yourself, because there's nothing there to find.

In 1988, Andrew Lo and Craig MacKinlay built a simple, clever test and pointed it at decades of U.S. stock data. Their finding, right there in the title: prices do not follow a pure random walk. The coin, it turned out, is very slightly bent. This is the story of how they caught it.

What "random walk" really claims

The random walk idea says price changes have no memory. Each day's move is independent of the last, like flipping a coin over and over. A run of heads tells you nothing about the next flip. Likewise, a few up-days should tell you nothing about tomorrow.

This was more than a theory, it was almost a moral position. If markets are efficient, all information is already baked into prices, so future moves must be unpredictable noise. Predictability would mean free money lying on the ground, and the argument went that traders would have snatched it up already. So: no memory, no patterns, pure randomness.

The question Lo and MacKinlay asked was refreshingly concrete: is that literally true in the data? Not "should it be true in theory," but "when we actually measure it, is it?"

The clever trick: measure risk at two speeds

Their tool is called the variance ratio test, and the idea behind it is genuinely neat. You don't need the formula, just the intuition.

Start with a fact about true coin-flips. If each day's move is an independent flip, then risk adds up in a very specific way as you stretch the time window. Two days of independent flips should be exactly twice as bouncy as one day. A week should be exactly five times as bouncy as a day. The bounciness scales in a clean, predictable, straight-line way. This is sometimes phrased as "variance grows in proportion to time," but you can just picture it as: for a memoryless coin, longer stretches are jumpier by an exact, known amount.

So here's the test. Measure how bouncy the market is over one day. Then measure how bouncy it is over, say, two weeks. Then check: does the two-week bounciness equal exactly what the one-day bounciness predicts for a memoryless coin?

• If yes, the ratio comes out to 1, the market behaves like a fair coin. No memory.
• If the longer-window bounciness is bigger than predicted, moves are reinforcing each other, up tends to follow up. That's momentum, a faint trend.
• If it's smaller than predicted, moves are partly cancelling, up tends to be followed by a pullback. That's mean reversion.

It's like checking whether a rolling snowball grows exactly as fast as pure geometry says it should. If it's growing faster, something is helping it along, there's structure, not just randomness.

What they actually found

Pointing this test at weekly U.S. stock returns, Lo and MacKinlay found the ratio was reliably bigger than 1, especially for broad indexes and portfolios of smaller companies. Short-horizon moves showed positive reinforcement, a mild tendency for the market to keep drifting in the same direction over the following days.

The random walk was rejected. Not smashed to pieces, the effect was small, not a neon sign, but statistically it was really there. The coin was bent.

Crucially, they were careful. A cheap way to find fake patterns is to lean on wild, once-in-a-decade days that distort simple measurements. Lo and MacKinlay designed their test to be robust to the fact that market volatility itself changes over time (the clustering that Engle's ARCH model describes). So their "the market has a little memory" conclusion wasn't an artifact of a few crazy weeks, it held up.

Why this mattered

This paper was a crack in a wall that had looked solid.

• It reopened the door to prediction. If prices have even a whisper of memory, then some patterns are real, and searching for them isn't automatically foolish. A huge amount of modern quantitative trading, momentum strategies especially, lives in the space this paper pried open.
• It sharpened the efficiency debate. "Markets are efficient" stopped being a blanket article of faith and became a measurable claim you could test horizon by horizon, market by market. The answer turned out to be "mostly, but not perfectly."
• The variance ratio test became a standard tool. Quants still use it today to probe whether a series trends, mean-reverts, or behaves like a coin-flip, a quick X-ray for hidden structure.

The honest limitations

Before you quit your job to trade the bend in the coin, the caveats matter, a lot:

• Small doesn't mean easy money. The departure from randomness was faint. Turning a faint statistical tendency into profit means trading a lot, and trading costs, fees, bid-ask spreads, slippage, can easily eat the entire tiny edge.
• The edge may have faded. Once a pattern is published, traders pile in and arbitrage it away. Effects that were measurable in older data are often much weaker now, precisely because papers like this one advertised them.
• A rejected random walk isn't a crystal ball. Showing the market has some memory is a long way from being able to reliably predict tomorrow. There's structure in the noise, but it's still mostly noise.
• Test enough things and something looks special by luck. Any hunt across many horizons and portfolios risks stumbling on a fluke. Lo and MacKinlay were rigorous, but the general warning applies to everyone who followed.

The one-line takeaway

Lo and MacKinlay caught the stock market cheating, just a little: it isn't a perfect memoryless coin-flip, short-term moves faintly reinforce each other, but the bend is so slight that seeing it is far easier than profiting from it.