nyasekai

Graveyard / Flows / ETFs

Trading ETF flows

Does a large inflow into a spot Bitcoin ETF, or a redemption shock in an equity ETF, predict returns?

Family
Flows
Verdict
Dead
Cause
No edge
Buried
9 October 2026

The idea

Money flowing into a spot Bitcoin ETF obliges the fund to buy Bitcoin, so flows tend to persist and push the price. For equity ETFs, an outflow shock may push prices below value and then revert.

How we tested it

DataTest 1: daily net flows of US spot Bitcoin ETFs (a public mirror of the Farside table). Buy IBIT at the open of the next day when the total flow is above $100M, hold while daily flow is positive. Variants: close entry, fading outflows, and an Ether ETF version. Test 2: long SPY and sector SPDR ETFs after redemption shocks (flow z-score below −2), from shares outstanding.
PeriodTest 1: 11 January 2024 to 8 October 2026 (2.7 years), split 22 March 2026. Test 2 not stated.
Costs1 bp per side; stress case 6.5 bp.
HoldoutIn-sample and out-of-sample gates, plus a pre-registered filter test.
ControlIBIT buy and hold over the same window (Sharpe 0.63).

Result

Bitcoin flow trades83
Sharpe full / in-sample / out-of-sample1.00 / 0.799 / 1.93
Out-of-sample trades18 (20 required)
Max drawdown / profit factor−24.96% / 1.76
Sharpe in bear markets−0.92
Fade variant, Sharpe0.58
Ether ETF version, Sharpe−0.29
Equity version, SPY / sector Sharpe−0.11 / 0.53 (0.50 in-sample, 0.72 out)

Why it failed

The Bitcoin version misses the in-sample Sharpe bar (0.799 against 0.8) and the out-of-sample trade count, loses in bear markets, and the volatility filter halves it. The difference in next-day return between being in and out of the position is only t of 1.2, so most of the result is being long Bitcoin about half the time in a rising market. The Ether version is negative. Equity ETF flow signals were flat or negative.

What would make it worth retesting

A re-run in 2027 with at least 30 out-of-sample trades and a regime filter that suits Bitcoin, defined in advance (a stock volatility filter is the wrong tool). A concentrated sector version needs a fresh pre-registration.

Past results of a backtest say nothing certain about the future.