Intraday Momentum: The First Half-Hour Predicts the Last

Most of a trading day's price action looks like noise once you've stripped out the overnight gap. But one 30-minute window doesn't: the S&P 500's return from the open to 10:00am has, across a long U.S. sample, a statistically reliable positive relationship with its return from 3:30pm to the close. Know the first half hour, and you have real information about the last half hour, hours before it happens. That's an unusually clean anomaly for something as picked-over as index-level returns, which is exactly why it's worth understanding both how it works and why it doesn't pay what the raw number suggests.

The shape of a trading day

Think of the trading day as having a memory that's short almost everywhere except at its own bookends. The middle of the day (late morning through early afternoon) tends to be closer to a random walk — dealers have absorbed the overnight news, volume is thin, and there's little fresh information forcing a direction. But the first half hour carries the market's initial digestion of everything that happened overnight: earnings, macro data, futures positioning. And the last half hour carries closing auctions, index-fund rebalancing flows, and the unwind of intraday hedges that were put on in response to that same morning move. The two ends are connected by the same information; the middle is comparatively quiet.

What the pattern looks like as a number

Gao, Han, Li and Zhou (2018) regressed the S&P 500's last-half-hour return on its first-half-hour return, across 30-minute bins from 1993–2013:

$r_{\text{last 30}} = \alpha + \beta \, r_{\text{first 30}} + \varepsilon$

In words: take the index's return in the closing half hour, and explain it with a straight line fit against that same day's opening half-hour return. They found $\beta$ reliably positive and statistically significant, with the relationship holding across most of their sample and strongest on high-volume, high-volatility days — days that are more likely to be driven by real information rather than noise.

Worked example. Suppose on a given morning the S&P 500 rises 0.30% in the first 30 minutes. With a fitted $\beta$ around 0.5 (in the range the original paper estimated for its strongest specifications) and $\alpha$ near zero, the model's expected last-30-minute return is $0.5 \times 0.30\% = 0.15\%$. A trader who buys index futures at 10:00am and holds to the close is betting on capturing roughly that 0.15%, scaled by position size — on a $10m notional futures position, about $15,000 of expected edge, before costs.

Drag points on this fitted line to get a feel for what the paper actually found: a real but noisy positive slope, not a tight relationship. Most of the scatter in the true data is unexplained — the last half hour still has plenty of its own randomness layered on top of the tilt from the first half hour.

Worked example, the mirror case. Now suppose the market opens weak: the first 30 minutes return $-0.40\%$. With the same $\beta \approx 0.5$, the expected last-30-minute return is $0.5 \times (-0.40\%) = -0.20\%$ — the model calls for a short. On a $10m notional futures position that's roughly $20,000 of expected edge, entered at 10:00am and held until the close. The sign of the trade flips with the sign of the morning move; the mechanism doesn't care whether the flow is buying or selling, only that it's correlated across both ends of the day.

Where the edge actually comes from — and what it costs to collect

The mechanism practitioners point to is order flow from large institutions who split trades across the day: an institution buying a large block often front-loads part of the order at the open (to reduce the risk of price drifting away from it) and executes another chunk near the close (to benchmark against the closing auction). That correlated buying at both ends of the day, driven by the same underlying institutional order, is what links the first half hour's direction to the last half hour's.

The catch is the holding period. Capturing this signal means being long or short the index from 10:00am to 4:00pm — nearly the entire trading day — to harvest a return that's a small fraction of a percent. That's very different from a fast signal you can exit in minutes. Over that six-hour window, the position is exposed to every intervening macro headline, earnings surprise, and liquidity shock the middle of the day happens to produce, and those risks are large relative to the 10–15bp of expected edge. The Sharpe ratio on the raw signal looks good in a backtest that ignores this; it looks much worse once you size the position for the six-hour value-at-risk it's actually carrying.

In interviews

State the empirical fact precisely — first-30-minute return predicts last-30-minute return, not the whole day — and give the plausible mechanism (correlated institutional flow at both bookends of the trading day). Then immediately raise the holding-period problem: any signal that requires holding through the entire middle of the day is competing against six hours of unrelated variance for a few basis points of edge, so the honest next question is always "what's the Sharpe, not the hit rate." If pushed on capacity, note that this is an index-level, not a stock-picking, signal — it scales with index futures liquidity, which is deep, but crowds fast once known, since everyone is trying to trade the same 30-minute window on the same instrument.