Intraday Volatility Forecasts for Execution Scheduling

An execution algorithm trading a large order over the day has to decide how much to trade in each interval, and that decision depends heavily on how risky the rest of the day is expected to be. Trade too much during a calm stretch and too little before a volatile close, and the order ends up exposed exactly when the market is most likely to move against it. The historical average volatility shape for a stock — typically a U-shape, high near the open and close, quieter midday — is a reasonable starting point, but it's not a forecast; today might not follow yesterday's average pattern at all, especially around news or events.

Average shape versus conditional forecast

The historical average intraday volatility curve answers "what does a typical day look like," aggregated across many past days. What a scheduling algorithm actually needs, though, is a conditional forecast: given what has happened so far today — realized volatility in the last hour, whether a scheduled economic release is imminent, whether the morning has already been unusually volatile — what does the rest of today look like? A model that only uses the unconditional average shape will be systematically wrong on exactly the days that matter most: unusually calm or unusually turbulent ones, which is when getting the schedule right or wrong has the biggest cost impact.

Building the forecast

A practical approach combines the historical average shape as a baseline with a short-horizon update based on recent realized volatility, similar in spirit to how a GARCH-type model updates a volatility forecast using recent squared returns rather than only a long-run average. Features include realized volatility over the last 15–60 minutes, the current bid-ask spread (which tends to widen alongside volatility), the time remaining until any scheduled news release, and the day-of-week or event-calendar context (index rebalance days, opex, earnings). The output — a volatility forecast for each remaining interval of the day — feeds directly into the scheduling optimization that decides how much of the remaining order to trade in each interval.

Worked example: an early volatility spike reshapes the schedule

A stock's historical average shows volatility roughly flat from 11am to 2pm. Today, realized volatility in the 10–11am hour comes in at three times its typical level, coinciding with an unexpected news headline. A model conditioning on that recent realization forecasts elevated volatility persisting through midday rather than reverting immediately to the flat historical baseline — a reasonable forecast, since volatility shocks in practice tend to decay gradually rather than vanish instantly. Acting on this forecast, the scheduling algorithm shifts more of the remaining order into the early afternoon, when the model expects the elevated risk to have partially subsided, rather than sticking to the historical U-shape's flat midday allocation, which would have kept trading at the same modest pace right through the still-turbulent period.

What this means in practice

A good intraday volatility forecast is the input, not the output, of the actual scheduling decision — algorithms like implementation-shortfall schedules trade faster (accepting more market impact) when expected volatility is high, since the risk of price drift while waiting outweighs the extra impact cost, and trade slower when volatility is expected to be low. Getting this forecast wrong in either direction has a real cost: overreacting to noisy short-horizon volatility spikes causes needless schedule churn, while ignoring genuine regime shifts leaves large orders exposed through periods of real risk.