Skip to content

Lead-lag trading

Lead-lag trading uses earlier moves in one asset to forecast another. Separate genuine prediction from stale data, unavailable messages and untradeable midpoint returns.

Lead-lag trading uses earlier price changes or order-flow information in one asset or market to forecast later changes in a related asset, then trades the expected response. The practical question is whether the follower still offers an executable opportunity after the strategy receives and processes the leader’s information.

An equity-index future and a constituent stock are a possible research pair: price discovery may happen at different speeds. That is a hypothesis about information transmission, not a promise that every futures move predicts the stock. Unlike a conventional pairs trade, the signal need not be a price spread expected to revert. The central feature is a directional, time-ordered relationship.

What exactly leads?

For returns measured over equal, non-overlapping intervals, one descriptive measure is the lagged correlation
ρAB(k)=Corr⁡(rtA,rt+kB),k=1,2,….\rho_{AB}(k)=\operatorname{Corr}(r^A_t,r^B_{t+k}),\qquad k=1,2,\ldots.

Here tt indexes interval endpoints and positive kk means A’s return comes earlier. A relationship can be positive or negative; “leader” names the temporal ordering, not the sign. Comparing positive and negative lags helps distinguish A-leading-B from the reverse ordering. The interval length and price definition must be stated: a five-minute last-trade return answers a different question from a sequence of quote changes.

A correlation peak is a diagnostic, not yet a forecast that improves on B’s own history. Test whether A adds information beyond the chosen baseline and shared market drivers. A common announcement can reach both markets while producing different adjustment speeds; this does not establish that trading in A causes B to move.

At high frequency, records arrive irregularly. Carrying forward B’s old price until its next update can manufacture an apparent delay relative to an actively updating A. Huth and Abergel’s methodology uses overlapping observation intervals to address asynchronous sampling. Hoffmann, Rosenbaum and Yoshida study estimation of a time shift from non-synchronous observations under an explicit stochastic model. Neither approach turns a statistically estimated lag into a guaranteed trade.

Scenario 1: the event is early; the message is late

Consider two fictional markets with exactly aligned clocks, all times in seconds after one origin:
EventTime
A’s market event occurs0.100
A historical replay submits a response using that event0.150
B’s market price moves0.200
A’s message reaches this strategy locally0.250
A’s event precedes B’s move by 0.100 seconds. But A’s message reaches this strategy 0.050 seconds after B has moved. The replay’s 0.150-second response uses information it would not yet possess: look-ahead bias. Sorting messages by exchange event time does not make them available at that time.
Fictional aligned-clock timeline: A event at 0.100 seconds, invalid replay decision at 0.150, B move at 0.200, and A local receipt at 0.250. B moves before this strategy receives A’s message.
Fictional aligned-clock timeline: A event at 0.100 seconds, invalid replay decision at 0.150, B move at 0.200, and A local receipt at 0.250. B moves before this strategy receives A’s message.

Original fictional timing scenario, with exact aligned clocks for illustration. The shaded interval begins at local receipt; processing and order-routing delays would push feasible action later still.

Keep four concepts separate: event time describes the source event; receipt time describes delivery to the specified system; decision time includes the processing needed to consume it; clock uncertainty limits how confidently timestamps from different systems can be ordered. An order must then reach its destination before it can fill. Receiving a signal does not establish that the old quote remains available.

The illustration assumes perfect clock alignment only to isolate the receipt problem. In real cross-market data, timestamp resolution and clock-offset uncertainty can be comparable to the measured lag. If the uncertainty intervals do not establish an ordering, report it as unresolved. A precisely printed timestamp is not evidence of an equally accurate clock.

The timestamp lesson develops this chain from input availability to fills. Its preview is public; the full lesson requires membership.

Scenario 2: a forecast is timely, but the midpoint is not a fill

This is a separate fictional trade, not an execution of Scenario 1. Assume the signal arrives in time and one B share can actually be bought for USD 100.20 at the ask and later sold for USD 100.80 at the bid. Assume both fills complete, with no additional price movement caused by the order. The pre-fee cash profit and return on the purchase outlay are
Profit=100.80−100.20=0.60 USD,Rbefore fees=0.60100.20≈0.5988024%.\begin{aligned} \text{Profit} &=100.80-100.20=0.60\ \text{USD},\\ R_{\text{before fees}}&=\frac{0.60}{100.20} \approx0.5988024\%. \end{aligned}
Suppose the corresponding midpoint prices were 100 and 101. Their 1% rise is a descriptive price change, not this trade’s return. The bid-ask spread is already reflected in the stipulated purchase and sale prices; do not subtract it a second time. Fees and other charges still reduce the USD 0.60 profit. A different order size, unavailable depth or delayed entry can change the fills themselves.
This distinction is empirically material. In their historical sample, Huth and Abergel’s forecasting backtest, §3.6, found predictive information in the leader but losses for their simple market-order strategy once the spread was included. That finding concerns their data and rule; it does not show that all lead-lag strategies lose, or that the same effect remains available today.

Evidence that survives a replay

A useful research design must challenge both the measured relationship and the trading policy:

  • Audit the observations. Check timestamp meanings, sequence gaps, stale quotes, corrections and overlapping trading hours. Compare trade-price and quote-based results: bid-ask bounce can contaminate trade returns even without a change in underlying value.
  • Preserve the information boundary. Replay receipt order and processing delays; start the forecast’s tradeable outcome after feasible entry. A future endpoint alone does not make a return interval eligible if part of it occurred before entry.
  • Freeze the research choices. Choose assets, lag windows and thresholds on training data, then evaluate in later periods. Trying many pairs and keeping the strongest peak invites overfitting. Check stability across sessions and news conditions rather than treating one fitted lag as permanent.
  • Reconcile positions with fills. Model entry and exit liquidity, fees, transaction costs, inventory exposure and any hedge. A follower-only position remains exposed to its own news; a cross-market hedge adds basis, funding and execution risks. Passive orders need a fill model, not an assumption of free spread capture.

Lead-lag evidence becomes economically useful only when information arrives early enough, prediction persists outside the fitting sample, and feasible trades retain value after costs. Temporal association by itself establishes none of those conditions.

Sources

Last updated
  • lead lag
  • cross market
  • timestamps
  • microstructure
  • execution
  • backtesting
Keep connecting the ideas
← Back to the glossaryBack to top ↑