The Hurst Exponent: Mean Reversion, Random Walk, or Trend - What H Actually Measures

TL;DR - The Hurst exponent H measures whether a return series remembers itself: H < 0.5 means mean-reverting (an up-move makes the next move more likely down), H = 0.5 means a random walk (no memory at all), H > 0.5 means trending (moves feed on themselves). We computed it two ways - classic rescaled-range (R/S) and detrended fluctuation analysis (DFA) - on eight years of daily log-returns (July 2018 - July 2026, 2,008 observations per name) for SPY, QQQ, NVDA, TSLA, VIX, GLD and USO, with a shuffle test as the null. The honest result: only VIX (H = 0.353, z = -4.2) and gold (0.405, z = -2.8) are genuinely different from a random walk - both mean-reverting. Not one instrument in the sample trends at the daily horizon. NVDA - the most-trended stock of the decade by eyeball - prints 0.539, statistically indistinguishable from coin flips. Two more findings matter for anyone quoting H numbers: the classic R/S estimator ran about +0.10 hot versus DFA on every series we tested (its small-sample bias is most of the "H = 0.6, it trends!" claims you will read), and the static number hides the tradeable part - SPY's rolling 2-year H swung from 0.64 in the 2021 trend regime to 0.33 in the 2022 chop, 0.32 in the May 2025 tariff whipsaw, and prints 0.43 today: a mean-reversion tape, which is exactly the regime in which chasing breakouts quietly bleeds.

Every systematic trader meets the Hurst exponent eventually, usually in a blog post claiming some ticker "has H = 0.65 and therefore trends." The number has real content - it is one of the few model-free ways to ask does this series have memory? - but most of what circulates about it fails the two tests any desk statistic must pass: is it estimated honestly, and does it survive a null. This post is the field guide we wished existed: what H is, how to read it, what it actually prints on real US market data, and where the estimator lies to you.

Interpretation first: H < 0.5 mean-reverting, H = 0.5 random walk, H > 0.5 trending

The Hurst exponent describes how the range of a cumulative return series grows with the observation window. Walk a truly random series for four times as long and its excursions grow by a factor of two - range scales like the square root of time, and on a log-log plot that square root is a slope of exactly 0.5. A series with H above 0.5 spreads out faster than chance: shocks persist, up begets up - trending, or in the statistical vocabulary, persistent / long-range dependent. A series with H below 0.5 spreads out slower than chance: excursions keep getting pulled back - mean-reverting, anti-persistent. The scale is not symmetric in practice: values below ~0.45 or above ~0.55 are rare in liquid markets, because anything further from 0.5 is a standing invitation to arbitrage.

Harold Edwin Hurst derived the statistic in 1951 from six decades of Nile flood records - the river's wet and dry years clustered far more than chance allowed (H ≈ 0.7), which is why the Aswan reservoir had to be sized for runs, not averages. Mandelbrot imported it into finance in the 1960s. The mechanics have not changed; the discipline of estimating it mostly has not arrived.

What we measured, and how

Eight years of daily log-returns - July 10, 2018 through July 8, 2026, 2,008 observations per instrument - for SPY, QQQ, NVDA, TSLA, the VIX index, GLD and USO: index beta, single-name momentum, implied volatility, and two commodities. Two independent estimators per series. Rescaled range (R/S): split the returns into blocks of length n, compute each block's cumulative-excursion range divided by its standard deviation, average, repeat across n from 8 to ~500 days, and regress log R/S on log n - the slope is H. Detrended fluctuation analysis (DFA): integrate the demeaned returns, remove a linear fit inside each block, measure the residual fluctuation F(n), same log-log regression. DFA is the workhorse of the physics literature precisely because it is less fooled by short samples and slow drifts.

The null is the part most quoted H values skip. An estimator on 2,008 points does not return 0.500 on truly random data - it returns something near 0.5 with sampling error. So for every instrument we shuffled its own returns into random order 200 times - identical distribution, fat tails and all, memory destroyed - and re-ran DFA on each shuffle. That yields the honest noise band: mean 0.50, standard deviation 0.035, 95% interval roughly 0.43 to 0.575. Any measured H inside that band is a random walk as far as this sample size can tell.

Instrument	R/S	DFA	z vs. shuffle null	Verdict
VIX	0.464	0.353	-4.2	genuinely mean-reverting
GLD	0.563	0.405	-2.8	mean-reverting
SPY	0.560	0.451	-1.4	random walk (lean reverting)
QQQ	0.560	0.462	-1.1	random walk
TSLA	0.603	0.524	+0.7	random walk
USO	0.587	0.528	+1.0	random walk
NVDA	0.589	0.539	+1.2	random walk

Two things in that table are worth more than the headline numbers. First, nothing trends. The most spectacular price trends of the sample - NVDA's decade, TSLA's manias - do not show up as daily-return persistence; a stock can multiply twenty-fold while its day-to-day increments stay statistically memoryless, because the trend lives in the drift, not in the autocorrelation H measures. Second, VIX is the one loud signal - H = 0.353, more than four standard deviations below the null. Implied volatility is structurally mean-reverting, which every options desk prices and which is why volatility spikes are sold, not chased. Gold's 0.405 is the quieter surprise, consistent with the choppy, headline-driven tape it printed over these eight years.

The estimator trap: R/S runs hot

Read the R/S column against the DFA column: R/S is higher for all seven instruments, by +0.05 to +0.16, averaging about +0.10. That is not a market fact; it is the well-documented small-sample bias of the rescaled-range statistic (the expected R/S of even white noise sits above the H = 0.5 asymptote until n gets large - the Anis-Lloyd correction exists precisely for this). By R/S alone, five of our seven series would be declared "trending" at face value. By DFA with a proper null, zero are. If a claim about some instrument's Hurst exponent does not name the estimator and does not come with a null, discard it - it is more likely measuring the estimator than the market.

The log-log picture makes the whole construction concrete: the slope of the fluctuation line is H. VIX's line grows visibly slower than the same returns shuffled into random order - that gap is the mean reversion, an effect you can see with a ruler. SPY hugs its shuffle line, which is the visual form of "statistically a random walk."

The static number is the least useful part

A single full-sample H answers a question nobody trades: was this series mean-reverting on average over eight years? The desk question is what is it now? So we rolled a 2-year (504 trading day) DFA window across SPY daily returns, stepping monthly:

The swings are the story. Through the post-COVID melt-up the rolling H reached 0.638 (May 2021) - a genuine persistence regime, the stretch when buying strength and holding winners worked and every dip-buyer got paid. It collapsed to 0.327 by September 2022, deep in the bear-market chop where every rally failed and every breakdown bounced. It bottomed again at 0.320 in May 2025 during the tariff whipsaw, and the latest window prints 0.427 - below 0.5, a tape that leans mean-reverting right now. Each estimate carries roughly ±0.07 of sampling error, so read the excursions and the regime, not the monthly wiggles.

That current reading is consistent with everything else we have measured on this tape recently: strong opens still fade at every gap size, earnings gaps give back their first hour more often than they extend it, and opening-range breakouts only pay under specific auction conditions. A sub-0.5 regime is a market that punishes chasing - the Hurst lens and the event studies agree.

One more cut, because prop traders live intraday: we ran both estimators on two years of 30-minute regular-session consolidated-tape bars (July 2024 - July 2026, 6,551 observations each). SPY: DFA 0.500. NVDA: 0.497. At the half-hour scale the tape is, by this measure, a coin toss - whatever intraday edges exist (and our event studies keep finding them) live in conditional structure around opens, auctions and news, not in unconditional bar-to-bar memory a Hurst exponent could harvest.

The honest limits

• H has no direction. A mean-reverting VIX tells you spikes decay; it does not tell you when, from what level, or how far. H is a regime dial, not a signal.
• It is slow. A stable estimate needs years of data, so by construction it describes the recent past. The rolling chart is a regime confirmation tool - position sizing and strategy selection - not an entry trigger.
• The error bars are real. ±0.07 on a 2-year window means a print of 0.46 and a print of 0.53 are the same number. Anyone quoting H to three decimals without a null is decorating.
• Estimator choice moves the answer more than the market does. R/S vs. DFA differed by more (+0.10) than any instrument in our sample differed from 0.5 (max -0.15). Name the method or the number is meaningless.

The desk rule, stated plainly: use H the way you'd use a weather report, not a forecast. Below ~0.45 on the rolling window, favor fading extensions and mean-reversion structures and make breakout systems earn their place; near 0.5, assume no memory and trade only conditional events; above ~0.55 - rare, and the 2021 window is what it looks like - let winners run longer than feels comfortable. On today's tape, the dial reads 0.43. The market is telling you what regime you are in. The Hurst exponent is just a disciplined way of listening.