Field Note 12 / Klines

Volume Predicts the Future. The Future Just Always Arrives

Does volume move before price? Yes, clearly. Does that fact help you catch a large move early? After three rounds of testing on Klines' own data, the honest answer is no, and the way it fails is more interesting than a plain “no” would be.

Round one: volume leads price by hours

The first test aligned every large rise (30–100%, n=2,217) to its own starting point and compared the hours before it to the coin's own background (median over the [−48h, −24h] window) and to a random-bar control (n=2,433). Six hours before a slow rise starts, volume sits 47% above background. Price range is only 12% above background at the same point. The control stays flat the entire time. That's a real, methodology-independent fact: volume moves first, and it keeps moving for hours while price barely does.

The honest caveat came in the same note: this shows what the average lead-up looks like, not what share of elevated-volume moments actually lead to a big move. A rule like “volume above 2x for six hours” would catch 38% of large rises, and it would also fire on 22% of ordinary random moments. As a standalone filter, useless. As a component, worth checking further.

Round two: the control was flattering itself

That further check found the first result was quietly cheating. A large rise, by construction of how it's labeled, starts at a local price minimum. Random bars aren't local minima. So features like “how far below the range” or “how negative the last hour's return” looked like powerful predictors (AUC as high as 0.91) purely because they're describing the definition of a pivot, not anything that precedes one.

Fixing this meant comparing large rises (≥30%, n=2,275) against a matched control of other local minima from the same labeling rule, ones that only produced a 5–10% bounce (n=5,981). Both are pivots by the same definition; the comparison becomes about what actually distinguishes them. Under that fairer test, the volume features from round one (15-minute volume, 15-minute and 5-minute trade counts) did not survive a confirmation split. Elevated volume before a big move is real. It's just equally present before a small one. It's a property of sitting at a local minimum, not a property of the minimum that's about to turn into something large.

Round three: does the order of activity matter, checked properly

One question was still open: not whether volume alone predicts anything, but whether the sequence in which volume, trade count, and price range each cross their own elevated threshold carries information, checked across every hour of continuous history rather than only around labeled events.

The test ran on 3,334,152 hourly grid points across 523 coins, November 2025 through September 2026 (one cascade excluded), with a baseline target rate of 2.45% for “currently inside a wave of 30% or more, less than half played out.” Two ways of defining “elevated” were used: a fixed 1.5x threshold (the level already measured in round one), and each coin's own rolling 90th percentile. The fixed threshold turned out to be too loose on its own: at 1.5x, 51.5% of all points share the exact same activity order, which measures nothing about the market and everything about the threshold being nearly always crossed.

With the stricter, self-calibrating threshold, sixteen possible orderings of volume, trades, and range were tested against a control matched by calendar day and turnover decile, with confidence intervals from a cluster bootstrap by coin (coins aren't independent; the ten busiest days alone produce 17% of all signals instead of the 3% you'd expect if activity were spread evenly). After correcting for testing sixteen orderings at once, zero of them passed at a 5% false discovery rate. Zero of eleven with the fixed threshold either.

The most common ordering (volume, then trades, then range) shows up on 31.9% of all points in the entire dataset and covers 67.6% of large waves. It also covers 68.5% of ordinary background moments at essentially the same false-positive rate (97.4% versus a population baseline of 97.55%), and the median outcome that follows it (a 6.3% favorable move, a −6.9% adverse one) is statistically indistinguishable from the market's unconditional background. The pattern is stable across five separate market regimes with base rates that vary threefold, which rules out one regime quietly driving the whole result. It shows up everywhere, catches most of what you want, and tells you nothing once it does.

What this actually establishes

This result is scoped deliberately: three states, a 12-hour window, two specific ways of defining “elevated,” on Binance spot 5-minute data across one bear-market year. It doesn't say volume-based structure never matters anywhere. It says that for this specific, honestly tested construction, the coverage you get is coverage you'd get for free, and it doesn't transfer into information about what happens next. Volume genuinely does lead price by hours. It leads price by hours before almost everything, which is exactly why it can't tell you which almost-everything is about to become something large.

Background on the detector this work sits on top of, and its own open questions: What Is Left for a Retail Trader? and the Klines lab page. The same heavy-tail problem this project keeps running into, where a handful of outcomes carry most of the signal and the rest is noise, is discussed from a different angle in Why Move Size Can't Be Predicted.

Author: Oleg Arefev · Karaptic Research · ORCID ↗ · SSRN author page ↗ · GitHub ↗