Why Order Flow Is So Autocorrelated

If you label every market order as +1 (buy) or −1 (sell) and look at the correlation between the sign of trade $n$ and trade $n+k$, you find something odd: it stays positive for thousands of trades, decaying only very slowly. Naively this looks like it should be arbitrageable — if buys predict more buys, why doesn't that just get priced in instantly? The resolution is the order-splitting hypothesis: what looks like a long run of independent bullish decisions is usually a single large decision, sliced into hundreds of small child orders and fed out over minutes or hours.

A worked illustration

A portfolio manager decides to buy 200,000 shares. An execution algorithm slices this into roughly 400 child orders of 500 shares each, released over two hours, interleaved with the market's normal background flow. Someone watching only the print tape sees a long, noisy-looking sequence: buy, buy, sell, buy, buy, buy, sell, buy... The buys aren't independent decisions — the same underlying 200,000-share order is generating dozens of the prints, so of course a buy print raises the odds the next print is also a buy: it might literally be the next child order from the identical parent.

Contrast this with what a naive model of "trader sentiment" would predict: that each trade reflects an independent new opinion, so sign correlation should vanish almost immediately (within a handful of trades) as opinions mix. Real markets don't look like that — the decay in autocorrelation is very slow, often taking hundreds of trades to fade, which matches the timescale of a large parent order being worked, not the timescale of individual sentiment.

Why it matters. Prices still stay roughly efficient despite this, because market makers and other liquidity providers widen or skew quotes in response to a run of same-sign flow, offsetting the impact of an order they correctly infer is still being worked — this balancing act is central to propagator models of impact. For anyone building execution algorithms, splitting is also the reason randomising the size and timing of child orders matters: a too-regular slicing pattern is itself a detectable signal that other participants can trade against, feeding Information Leakage and Signalling Risk.