Heteroskedasticity: Why Model Errors Change Size in Markets

A backtest hands you a clean-looking number. The strategy’s average edge is 0.4% a week, the confidence interval sits comfortably away from zero, and the significance test says a result like this would show up by chance less than one time in twenty. On that evidence you’d size a position. The problem is that every one of those figures rests on a quiet assumption about the model’s errors, and it’s an assumption markets break more often than they keep.

That assumption has a name. Heteroskedasticity is the situation where the spread of a model’s errors is not constant across observations. When it’s present and you ignore it, the line through your data can still look reasonable while the uncertainty you’ve attached to it is wrong. This piece is about spotting that gap and reading confidence claims with the right amount of doubt.

What heteroskedasticity actually describes

Start by separating two things that get blurred together. A return series is what an asset actually did: a column of weekly or daily percentage moves. A model residual is something else. It’s what’s left over after you fit a relationship. Run a linear regression of a stock’s return on the market’s return, and for each week you get a modelled return and an actual return. The residual is the gap between them.

Heteroskedasticity describes the behaviour of those gaps. Under the tidy assumption most tests lean on, the residuals have roughly the same spread everywhere, whatever the market was doing. That is homoskedasticity, constant error variance. Heteroskedasticity is the case where the residuals are tightly clustered in one period and widely dispersed in another. The average relationship the line describes can be perfectly sensible while the errors around it breathe in and out.

The distinction matters more than it sounds. A model can have a good average fit and still be heteroskedastic, because the two are separate questions. One asks whether the line sits in about the right place. The other asks whether the scatter around the line is stable enough to trust the error bars you print underneath it.

Why the errors change size in market data

Markets produce changing error variance for reasons you already know from the chart. Volatility clusters: calm weeks follow calm weeks, and violent days arrive in bunches. A model calibrated across both regimes will hug the data when things are quiet and miss by a wide margin when they aren’t. The residual spread simply inherits that clustering.

Event periods do the same thing on a schedule. Around an earnings release or a rate decision, the gap between what a model expects and what price delivers can blow out for a few sessions and then settle. Liquidity shifts widen it too, since thinner books move further on the same order flow, so the unexplained part of a return grows. Price-level effects appear when you model raw prices rather than returns, because a fixed percentage move is a larger absolute move at a higher price. And a structural break, a genuine change in how the series behaves, can shift the error variance permanently rather than for a few sessions.

When I plot residuals against time, the shape I’m hunting for is a fan: errors that sit close to the line through quiet months and flare out around the noisy ones. On one market model I kept for a large-cap name, the residual band ran near plus or minus 0.8% through a calm quarter and pushed past 3% in a single event week. Same model, same coefficients, roughly four times the scatter.

Where it quietly breaks a conclusion

The coefficient estimates from a regression are often still usable when the errors are heteroskedastic. The damage lands on the uncertainty around them. Standard errors, confidence intervals, t-statistics and p-values are all computed under the assumption of constant error variance. Break that assumption and those figures rest on a footing that no longer holds.

The direction of the error is what makes it dangerous. Ignoring heteroskedasticity usually makes an estimate look more precise than it is. The confidence interval prints too narrow, the t-statistic too large, and a relationship crosses into significance when a fair accounting would leave it short. You can measure the standard deviation of a series carefully and still hand yourself a false read on precision, because the problem lives in how the error spread moves, not in any single summary number.

Here is the first misread to guard against. A tight confidence interval says nothing on its own about the strength of an effect. It can just as easily be an artefact of an assumption the data quietly violated. When a backtest reports a suspiciously clean significance level on noisy market data, changing error variance is one of the first things worth suspecting.

A two-period example worth sitting with

Picture a model run over two six-month windows. In the calm window the average absolute error is 1.2%. In the turbulent window it’s 1.3%. On the averages alone you’d call the model equally reliable in both, and you’d have missed what matters.

Look inside the windows instead of at their means. The calm window keeps almost every residual inside a 2% band. The turbulent window has the same central tendency but swings from near zero to 5%, with the biggest misses stacked into a handful of volatile weeks. The average quietly absorbed those extremes and hid them.

The right reading is about precision, not prediction. In the turbulent window the model’s output deserves a wider band of doubt, so any estimate drawn from it should carry more uncertainty than the tidy interval the calm window earned. None of this tells you which way price goes next. It tells you how much to trust the confidence figure attached to the model, and the honest answer changes between the two periods even though the average error barely moved.

How I check for it

The first check is the cheapest one, and it’s visual. Plot the residuals against the fitted values, then against time. Constant variance looks like a shapeless cloud of even thickness. Heteroskedasticity looks like a fan, a funnel, or a band that clearly thickens in some stretches. Your eye catches it before any formal test does.

The second check splits the sample. Take a calm subsample and a turbulent one, measure the residual spread in each, and compare. If the turbulent-period spread runs several times the calm-period spread, the constant-variance assumption is already gone and no test is needed to confirm it. I lean on this comparison more than any single statistic, because it maps straight onto the volatility regimes I can already see on the chart.

The third step is the fix people reach for: heteroskedasticity-robust standard errors. These recompute the uncertainty around your estimates so the error bars widen to reflect the changing variance. Reported precision comes down, significance levels get more honest, and a t-statistic that was flattered by the constant-variance assumption comes back to earth. That is a real improvement in the one thing it touches.

What robust standard errors will not do

Here is where the second misread hides, and it’s a costly one. The correction changes the reported uncertainty around some estimates. It does nothing else. It will not repair a poorly specified model, it will not recover a variable you left out, it will not turn correlation into causation, and it will not make an unstable relationship stable. A model that’s wrong about the average relationship stays wrong after you apply it. You’ve corrected the error bars, not the line.

That limit connects to a wider research hazard. Search enough specifications and you’ll find one that clears a significance threshold by luck, a problem laid out in the work on multiple-testing risk. A robust standard error offers no protection there. It gives you an honest confidence figure for the single model in front of you. It can’t audit how many models you tried before this one survived.

So treat the correction as an adjustment to one specific claim, the precision of an estimate, and nothing wider. When someone waves an honest-looking t-statistic as proof that a strategy works, the useful response is to ask about the model, the omitted variables, and the number of specifications tested before this one came through.

The limitation to keep in front of you

Heteroskedasticity is model-dependent, and that’s the point to leave on. You can’t detect changing residual variance without first stating a model, because there are no residuals until something has been fitted. Change the model and the residuals change with it. What looks like unstable error variance under one specification can shrink under a better one that captures the structure you were missing. Adequate data matters too: a handful of observations in the turbulent window won’t tell you much, and no single diagnostic settles the question on its own.

The traders I learn the most from treat every confidence number as a claim to be interrogated rather than a fact to be accepted, and Nassim Taleb built a career on that suspicion of tidy statistics drawn from wild data. His lessons on fat tails and fragile estimates sit well beside this idea, because changing variance is one of the concrete mechanisms behind the fragility he warns about.

Carry one habit out of this. Before you trust a confidence interval or a significance test on market data, ask whether the errors behind it hold a constant spread, and check rather than assume. That single question separates a number you can lean on from one that’s flattering you. Learn the pattern. Ride the trend. Keep the gains.

Educational content only. Not investment advice. Trading involves risk. You are responsible for your decisions.