1. One beta is an average of two very different regimes
Ordinary market beta is the slope of a regression of an asset's returns on the market's — a single coefficient fit over every day in the sample. It is the model-free summary of systematic risk that the whole CAPM machinery rests on. But it lumps two regimes into one line: the days the market rallied and the days it fell. An asset that drifts loosely with the market when things are calm but slams to it during a sell-off has an average beta that flatters it — the number you'd size a hedge on hides the behaviour that actually matters when the hedge is tested.
2. Splitting beta by the sign of the market
The fix is to stop fitting one line and fit two, each over the days the market moved a particular way. Ang, Chen & Xing (2006), in their Downside Risk study, define downside beta and upside beta as ordinary regression betas computed on the conditional subsamples:
β⁻ = Cov(rᵢ, r_m | r_m < 0) / Var(r_m | r_m < 0)
β⁺ = Cov(rᵢ, r_m | r_m > 0) / Var(r_m | r_m > 0)
Each is just a beta — a covariance over a variance — but estimated only over the days the market fell (β⁻) or rose (β⁺), with both the covariance and the variance measured relative to the conditional means. Nothing exotic: one linear regression per regime. What changes is the question. Instead of “how does this asset move with the market on average?” you are asking “how does it move with the market when the market is falling?”.
3. Why the downside half is the one that's priced
Investors are not indifferent between the two. A dollar lost in a market crash hurts more than a dollar missed in a rally feels good — the core intuition of loss aversion and downside-risk preferences. Ang, Chen & Xing show this empirically: stocks with high downside beta earn higher average returns, a premium that ordinary beta does not capture. In other words, the market pays you to hold assets that grip it on the way down, precisely because that co-movement is what ruins a portfolio when it can least afford it. Downside beta isolates the sensitivity that carries the risk premium; upside beta is the pleasant half that doesn't.
4. Beta asymmetry: the number to watch
The single most useful readout is the gap between the two:
beta asymmetry = β⁻ − β⁺
A positive asymmetry means the asset tightens its grip on the market exactly when the market falls — extra downside risk that an average beta of, say, 1.0 would never reveal. A name can look like a tame, market-neutral hold on its blended beta and still carry a large, positive β⁻ − β⁺: fine in the good times, a liability in the bad ones. Tracking the asymmetry turns “this hedge behaves differently in a crisis” from a war story into a monitorable figure.
5. Where it sits next to semicovariance
Downside beta is the single-name, market-conditioned cousin of the cross-asset decomposition we covered in realized semicovariance. Both isolate the downside co-movement that realized covariance and ordinary beta average away — but they condition differently. Semicovariance's negative component keeps the days where both assets fell; downside beta conditions on the market falling and reads the asset's sensitivity given that. And where realized semivariance splits a single asset's own variance by the sign of its return, downside beta splits its systematic exposure by the sign of the market. Same asymmetry principle, three complementary lenses on the risk that only shows up on the way down.
6. A caveat about conditioning
Conditioning on the market's sign is not free of statistical hazard. Selecting a subsample by an outcome changes its distribution, and conditioning on larger moves can mechanically inflate measured co-movement — the correlation-conditioning bias documented by Forbes & Rigobon (2002) and others. Signing (splitting at zero) is milder than conditioning on magnitude, but the lesson stands: read β⁻ against β⁺ and against the unconditional beta, not in isolation, and make sure each side has enough observations to estimate a slope at all. A downside beta computed from a handful of down days is noise wearing a decimal point — treat it as undefined until the sample earns it.
7. Computing it
Our open-source orderflow-metrics library ships all three, dependency-free, in TypeScript and Python:
import {
downsideBeta,
upsideBeta,
betaAsymmetry,
} from "orderflow-metrics";
// aligned asset and market returns
const market = [0.01, -0.02, 0.015, -0.01, -0.03, 0.02, -0.015, 0.005];
const asset = [0.008, -0.03, 0.012, -0.02, -0.05, 0.018, -0.02, 0.004];
downsideBeta(asset, market); // 1.6 — beta over the days the market fell
upsideBeta(asset, market); // 0.92 — beta over the days the market rose
betaAsymmetry(asset, market); // 0.68 — β⁻ − β⁺, the priced downside sensitivity
This asset looks roughly market-like on a blended beta, but it grips the market far harder on the way down (1.6) than up (0.92) — a 0.68 asymmetry a single number would have hidden. The Python distribution exposes the same functions (downside_beta, upside_beta, beta_asymmetry), returning NaN when a side has too few qualifying days to fit. They sit beside the realized covariance / correlation / beta and semicovariance tools inside the wider market-microstructure toolkit we build in the open — install it from our open-source page (npm and PyPI, MIT-licensed).
8. Conclusion
A single beta answers “how much does this move with the market?” and stops there. But the market does not price up-moves and down-moves the same way, and neither should your risk model. Downside beta measures the sensitivity that shows up in a sell-off, upside beta the one that doesn't, and their gap is the asymmetry investors demand a premium for — three regressions on one return series, and a tame-looking position reveals whether it is really tame or just tame on average. Explore the rest of the toolkit in our quantitative research library, or read the implementation in our open-source metrics.