Methodology › Research
Does crypto order flow predict prices? We tested our own data
Whalio publishes who is buying and who is selling: buy ratio, volume delta, net flow and single trades of $10,000 or more. The question we hear most is whether any of it can be traded. Rather than answer from intuition, we ran the test that a fund would run before putting money behind a signal, on our own data, and we are publishing the result whichever way it came out.
On direction it came out negative. Across 113 tests on 20 large coins, no signal beat trading costs in a way that held up. On size it came out positive: how calm or busy a coin has been says a good deal about how far it is likely to move next, on every coin we tested. That finding is what the market readout is now built on.
What we found, in short
- Extreme buying does not lead to higher prices. An hour in which a coin's buyers were more aggressive than in 95% of its previous month was followed by nothing at all, on spot and on futures.
- Some intuitive signals point the wrong way. A burst of large trades with buyers in front was followed by a slight fall, not a rise. Net buying into a hard drop was followed by more underperformance, not a rebound.
- The one robust pattern is too small to trade. Eight hours of net buying in a row came before a move of about 0.1% above the market over the next four hours. A round trip on Binance spot costs about 0.2%.
- One direction pattern deserves another look. Crowded longs, measured by funding, came before underperformance on 15 of 19 coins. It is not proven. A second candidate, large traders against the crowd, disappeared once both sides were measured in the same market.
- The size of the next move is predictable. After an unusually quiet day, the next day moved about 40% more than the quiet day suggested, on 20 of 20 coins. A forecast built on this held its 80% range 81% of the time on months it had never seen.
How the test was run
A backtest is easy to fool, mostly by accident. These are the rules we held to, and the reason for each.
- No look-ahead. A signal at a given hour uses only data that existed when that hour closed. Returns are measured from that close.
- Each coin against itself. “Extreme” means extreme for that coin over its own previous 30 days, the way the site ranks readings.
- The market is subtracted. Every result is shown as excess return over the average of all 20 coins in the same hours, so a signal gets no credit for a rising market.
- No double counting. Once a coin fires, it cannot fire again until the holding period is over.
- Days, not trades. Coins tend to fire together on the same day, so the t-statistics treat each day as one observation.
- Stable or not at all. Each result is checked separately in the first and second half of the period, and coin by coin.
- Costs. Binance spot charges about 0.10% a side to a taker, so a spot trade has to earn 20 basis points to break even. Futures cost about half that.
- Many tests, some flukes. At the usual bar of |t| above 2, about one test in twenty passes by luck. With 113 tests, about six lucky passes are expected.
The coins are BTC, ETH, BNB, XRP, SOL, TRX, ZEC, DOGE, LINK, ADA, XLM, LTC, UNI, HBAR, AVAX, SHIB, NEAR, TAO, AAVE and DOT, tested at 1, 4 and 24 hours. A basis point (bp) is 0.01%. The tables show selected rows; the counts above cover every test.
1. Spot order flow
The measurements the main dashboard is built on, from Binance spot candles: what share of each hour's volume hit the ask, and the running balance of buying and selling. 36 tests. 3 cleared |t| above 2, where 1.8 would by chance.
| Signal | Hold | Events | Excess, bp | t | Coins agree | Reading |
|---|---|---|---|---|---|---|
| 1-hour buy ratio at or above its 95th percentile | 1h | 8,662 | −0.4 | −0.80 | 9/20 | Nothing |
| 24-hour net flow at or above its 95th percentile | 24h | 1,059 | +4.6 | −0.12 | 11/20 | Nothing |
| 24-hour net flow at or below its 5th percentile | 24h | 970 | −8.1 | −0.88 | 12/20 | Nothing |
| Volume three times normal, buyers leading | 24h | 1,487 | +12.7 | +1.42 | 11/20 | Not significant |
| Price fell hard while buyers were net ahead | 24h | 462 | −15.2 | −1.27 | 9/20 | Points the wrong way |
| Eight or more hours of net buying in a row | 1h | 1,904 | +4.1 | +2.30 | 15/20 | Real but tiny; see below |
| Eight or more hours of net buying in a row | 4h | 796 | +10.4 | +1.90 | 14/20 | Loses money after fees |
The streak of net buying is the one spot result that looks alive. It is not robust: it works at 8 hours, fades at 6 and 10, and turns negative at 16. And a long trade on it held for four hours lost 7.2 bp on average after a 20 bp round trip, with 41% of trades profitable. At futures-level costs of 10 bp it would have made about 2.8 bp a trade, which is not enough to survive slippage.
2. Futures: flow, funding, premium and positioning
Most crypto volume trades on perpetual futures, about six times spot for these coins, so we ran the same test on Binance USDT-margined perpetuals: their buy ratio, the funding rate longs pay shorts, and the premium of the perpetual over spot. Open interest and long/short ratios are only published for the last 30 days, so those were tested on 25 September to 11 October 2026 alone. 53 tests. 1 cleared the bar, where 2.7 would by chance.
| Signal | Hold | Events | Excess, bp | t | Coins agree | Reading |
|---|---|---|---|---|---|---|
| Perpetual 1-hour buy ratio at or above its 95th percentile | 1h | 8,656 | +0.7 | +0.29 | 14/20 | Nothing |
| Perpetual 24-hour net flow at or above its 95th percentile | 24h | 1,121 | +11.9 | +1.95 | 12/20 | Borderline; second half only |
| Funding at or above its 95th percentile (crowded longs) | 24h | 480 | −27.9 | −1.76 | 15/19 | Closest to real; not significant |
| Funding at or below its 5th percentile (crowded shorts) | 24h | 741 | −1.0 | −1.45 | 10/19 | Nothing |
| Perpetual premium over spot at or above its 95th percentile | 24h | 2,727 | −6.1 | −0.78 | 13/20 | Nothing |
| Spot net buying while perpetuals net sell (24 hours) | 24h | 3,048 | −2.8 | −0.32 | 8/20 | Nothing |
| Most accounts long while top traders are short | 24h | 45 | −89.6 | −2.45 | 2/3 | 17 days, 45 events: too few |
Funding is the most interesting line. When longs were paying more than in 95% of the previous month, the coin trailed the market by 27.9 bp over the next day, in both halves of the year and on 15 of 19 coins. That is the shape a real effect has. It still falls short of the bar, at t of −1.76, so for now it is a hypothesis and not a finding. The only futures result that cleared the bar rests on 45 events in 17 days, which is too few to mean anything.
3. Whale trades
Whalio's Huge Trades feed records single trades of $10,000 or more. They are futures trades: hour by hour their net buying tracks Binance perpetual flow (correlation 0.52 on 19 of 19 coins) far more closely than spot flow (0.22). The history kept so far is short, 15 September to 11 October 2026, and most of it fell in a strong rally, which makes this the weakest of the tests. 24 tests. 3 cleared the bar, where 1.2 would by chance.
| Signal | Hold | Events | Excess, bp | t | Coins agree | Reading |
|---|---|---|---|---|---|---|
| Large-trade net buying in the hour at or above its 95th percentile | 1h | 640 | −5.6 | −1.79 | 5/19 | Points the wrong way |
| Large-trade count three times normal, buyers leading | 1h | 303 | −13.4 | −2.43 | 1/14 | Wrong way, one coin in 14 |
| Large trades net buying while spot net sells (24 hours) | 24h | 181 | +57.1 | +2.24 | 13/19 | Artefact: two markets mixed |
| Large trades net selling while spot net buys (24 hours) | 4h | 924 | −6.2 | −2.09 | 14/19 | Artefact: two markets mixed |
| Large trades net buying while the rest of the futures tape sells | 24h | 257 | +10.5 | +0.27 | 9/19 | Nothing, in one market |
| Large trades net selling while the rest of the futures tape buys | 24h | 169 | −9.0 | −0.63 | 11/17 | Nothing, in one market |
The first version of this study reported the third and fourth lines as its one coherent pattern: when large trades leaned one way and the rest of the market the other, the price followed the large trades. That comparison set futures trades against the spot tape. Measured within one market, large futures trades against the rest of the futures tape, the effect is gone, as the last two lines show. We have kept the original lines in the table so the correction is visible.
4. The size of the next move: where the data does work
Every test above asks which way the price goes. A different question is how far it goes, whichever way. Volatility clusters, so the fair baseline is not zero but what the last 24 hours already imply. Each line below shows how much bigger or smaller the next move was than that baseline, after the reading, over the same 11 months and 20 coins.
| Reading | Hold | Events | Move size | t | Coins agree | Twice as big |
|---|---|---|---|---|---|---|
| Last 24 hours calmer than 90% of the past month | 24h | 1,620 | +42% | +17.4 | 20/20 | 21.1% |
| Last 24 hours calmer than 90% of the past month | 4h | 5,829 | +34% | +19.0 | 20/20 | 14.1% |
| Last 24 hours wilder than 90% of the past month | 24h | 989 | −22% | −11.9 | 20/20 | 0.9% |
| Spot volume three times a normal hour | 1h | 15,013 | +13% | +6.1 | 19/20 | 10.2% |
| Futures volume three times a normal hour | 1h | 13,070 | +16% | +7.1 | 20/20 | 10.7% |
| Large-trade count three times normal (27 days) | 1h | 709 | +17% | +3.1 | 12/14 | 8.3% |
| Buy/sell imbalance in the top 5% | 1h | 8,418 | −3% | −1.8 | 11/20 | 6.5% |
| Funding far from its usual level | 24h | 1,446 | −3% | −1.3 | 16/20 | 6.4% |
“Twice as big”: the share of events after which the next move was more than double what the past day implied. Across all hours it is about 7%.
This is a different kind of result from everything above. The quiet-day effect holds on every one of the 20 coins, in both halves of the year, with t-statistics above 15. Calm periods are followed by larger moves than they suggest, and wild ones by smaller. Volume bursts add a little on top, mostly because they hit the whole market at once. Buy pressure and funding say nothing about size either.
We built a forecast on it: a standard model of realised volatility (HAR) using the coin's last day, week and month, its volume and the time of day. Fitted on November to May and judged on May to October, which it had never seen, it beat “tomorrow will look like today” by 24% at 24 hours and 19% at 4 hours. Its 80% range held 81.1% of the time at 24 hours and 80.1% at 4 hours; its 95% range held 95.0% and 95.6%; no coin fell outside 78–84% for the 80% range. It is published for every coin, hourly, on the market readout, where it is also re-checked every day against all tracked coins and against its own published record.
What this means if you use Whalio
Order flow describes a market well. It shows whether a move was bought or sold into, whether large traders took part, and whether a reading is unusual for that coin. On this evidence it does not, on its own, tell you where the price goes next, at least not by enough to pay for the trade.
It does tell you how far the price is likely to go, which matters for decisions that do not depend on direction: how large a position can be for a given risk, whether a stop sits inside ordinary noise, whether a quiet market is likely to stay quiet. That is what each coin page and the market readout now lead with, and every other reading there carries a label saying what this study found it is good for.
That is also a useful yardstick for anything sold as a signal built from the same kind of data. Ask to see this test: market-adjusted, non-overlapping, split in halves, and after costs.
Limits of this study
- Eleven months is one market regime, not a cycle. Effects that appear only in panics or manias would not show here.
- One exchange. Binance is the largest venue, but flow elsewhere is not in the data.
- Hourly resolution. Anything that plays out in minutes is invisible at this scale.
- Twenty large coins. Small caps behave differently, and their costs are higher.
- Simple rules, tested one at a time. A model combining several inputs might do better, or might just fit the noise better.
- The move-size forecast was fitted on 20 large coins. It was not fitted on smaller coins, which is why its hit rate is re-checked daily on every coin the site tracks.
This is research about data, not advice. It does not know your circumstances and does not tell you what to buy or sell. Past patterns, including the ones above, can stop working without warning. Crypto-asset prices are volatile and you can lose everything you put in.