ETF Flows: the Number the Whole Market Watches, and Why It Does Not Tell You Where the Price Goes
· Strategy
Since January 2024 every morning in the Bitcoin market starts with the same number: how much money went into the US spot BTC ETFs yesterday. "Record $1 bn inflow", "third day of outflows". It sounds like the cleanest possible signal of institutional demand, and like something that should lead the price.
We tested it the way we test every idea: not on a chart, but as a trading rule that can be run and measured. We collected the flow of every session since 11 January 2024, the minute-by-minute tape of IBIT and FBTC, and built a forecast of the day's flow before it is published. The result is simple and a little uncomfortable: flows describe what happened very well, and say almost nothing about what will happen.
1. The link is there, but on the same day
A session's flow and BTC's move on the same day are clearly related: a correlation of 0.41 over the whole period, and 0.61 in 2026. Move one day forward and the link is gone. The correlation of the flow with BTC's move over the next one, three and seven days is 0.03–0.05, practically zero.
This is not chance but mechanics. When the price rises, investors buy fund shares, the fund buys bitcoin, and in the evening we see a big inflow. The flow is a trace of the price move, not a cause that could be seen in advance.
2. We can forecast the flow before it is published
So we thought: if the link is on the same day, maybe it is enough to know the flow before the rest of the market does? The official figures come out in the evening or the next day, but the flow leaves a trace in the funds' own trading: their price against BTC and their volume during the session.
A model that reads the IBIT and FBTC minute bars every 30 minutes and learns from the previous 60 sessions gets the direction of the day's flow surprisingly right. On 202 sessions of 2026:
| Forecast time (New York) | Direction right | On days with the bigger flows (half of the sessions) | Correlation with the flow |
|---|---|---|---|
| 12:16, mid-session | 69% | 81% | 0.54 |
| 16:16, after the close | 73% | 82% | 0.65 |
In other words, hours before publication we know whether the day will be an inflow or an outflow day, three times out of four.
3. And it tells us nothing about the price
This is the most important lesson. We tested the forecast honestly: on 659 sessions of 2024–2026, computed only from data available at the time, and measured BTC's move only from the moment the forecast is known: one hour, four, eight, a day and three days on. In every horizon the correlation stays within ±0.1. Splitting the days into fifths from the strongest outflows to the strongest inflows gives no ordered pattern. Even the "surprise", the part of the flow the price move alone does not explain, predicts nothing.
The high same-day correlation builds up before the forecast is known. By the time we know it, the market has already priced it in.
4. The fund premium: an edge that evaporated
We also checked a subtler signal: the premium, how much more than usual the fund shares trade for against BTC during the session. In 2024 the days with the highest premium (top 10%) did come before better 24 hours for BTC. Then the effect faded:
| Year | BTC over 24 h after the highest premium | BTC on the other days |
|---|---|---|
| 2024 | +0.82% | +0.21% |
| 2025 | +0.18% | −0.06% |
| 2026 | −0.06% | −0.15% |
This is the fate of most simple edges in a liquid market: as more participants watch the same data, the difference wears away. With positions held for days and stops of 1–2%, an edge of 0.1–0.2 percentage points a day is smaller than the noise and the costs.
5. A test in the strategy: big outflows as a short signal
One result looked promising. After sessions with an outflow of more than $500 M, BTC lost 1.4% on average the next day and 1.7% over a week (25 such sessions). So we built a SHORT scenario triggered by a big outflow and ran it through a full backtest of our strategy on three runs with shifted entry times.
- The ETF-era result was worse on all three runs: by 19%, 54% and 66%. The scenario opened 9–14 trades with an average result close to zero, and meanwhile took the slot from better entries.
- The opposite idea, do not buy after a big outflow, looked like +12%. Comparing the trade lists showed the filter changed exactly one ETF-era trade: in the strategy's history no long ever entered after a session with an outflow above $500 M. The whole "gain" was an accident of there being no flow data before January 2024.
6. The worst trades did not go against the flow
One last approach: perhaps the flow forecast works as a safety catch, blocking entries when the flow goes against the position? We went through the strategy's 449 trades since March 2024. Most of the 25 worst were opened when the flow forecast agreed with the position's direction, not against it. Blocking entries against the forecast (a $200 M threshold) would have stopped 34 trades, every one of them a winner. An early exit at the first adverse forecast cut three times more winners than losers. Even the actual day's flow, known in hindsight, lowered the result when used as a block.
7. Where flows really helped: fresh data
Paradoxically, the most valuable use was the least spectacular one. Some of our scenarios read the total bitcoin held by the funds, which our on-chain data provider publishes 1–2 days late. So two thirds of the time the live trader saw a value from two days ago, while the backtest used yesterday's. We added an estimate: the last published holdings plus the day's flow converted at the closing price.
- The agreement of the scenario conditions with the data the provider finally publishes rose from 89.8% to 96.7%, with an estimate error of about 0.5%.
- A backtest on data a day late lost about 6% and two good shorts. With the estimate the strategy kept every trade, and the estimate's error added one needless entry.
For us flows are therefore not a signal but material that lets the on-chain data keep up with the market.
Summary
- ETF flows move with the price on the same day (correlation 0.41–0.61) but do not lead it: a day later the correlation falls to almost zero.
- The day's flow can be forecast from the funds' tape with about 70–75% accuracy, and yet from the moment the forecast is known it does not predict BTC's move.
- The edge from the fund premium was real in 2024 and had faded by 2026.
- Trading rules built on flows (a short after outflows, a long filter, entry blocks and early exits) made the strategy worse or changed nothing.
- Flows gave real value as supporting data: thanks to them the live trader sees the fund holdings the way the backtest assumes.
The general lesson goes beyond ETFs: a same-day correlation is not a forecast. Every indicator has to be tested from the moment it is actually known, and the market's most famous numbers are usually in the price before we get to read them.