Skip to content

Lead variable

A lead variable aligns a future observation with an earlier row. See how leads form forward-return targets, when they leak future data, and which future inputs can already be known.

A lead variable is a value of a time series aligned with an earlier observation. The hh-step lead at row tt is yt+hy_{t+h}, where hh is a positive integer. In a return-prediction dataset, a lead often supplies the future outcome the model is meant to predict. Supplying that realised outcome as a live input instead creates look-ahead bias.
The shift changes where a value sits in a table. It does not change when anyone could have known it.

Three prices, two completed labels

Take synthetic prices of 100, 103 and 101 USD per share at equally spaced observations t=0,1,2t=0,1,2. Each price becomes known at its own observation time. There are no distributions or corporate actions; returns below are simple price returns before costs.

Row timeCurrent price PtP_tOne-step lead Pt+1P_{t+1}Forward-return label Pt+1/Pt−1P_{t+1}/P_t-1
t=0t=01001033%3\%
t=1t=1103101−2/103≈−1.9417%-2/103\approx-1.9417\%
t=2t=2101MissingMissing
This is a retrospective table, assembled after observing the prices. At t=0t=0, neither the realised lead of 103 nor its derived return of 3%3\% is available. They become known at t=1t=1 under the stated immediate-availability convention. The final row has no observed successor; its missing lead is not zero and must not wrap around to the first price.
At decision time t0, price 100 USD per share is known; prices 103 at t1 and 101 at t2 are still future observations. Aligning 103 to row t0 gives the later-completed target and a 3 percent return label, but cannot make it a live feature at t0.
At decision time t0, price 100 USD per share is known; prices 103 at t1 and 101 at t2 are still future observations. Aligning 103 to row t0 gives the later-completed target and a 3 percent return label, but cannot make it a live feature at t0.
Original synthetic timeline viewed from t=0t=0. Filled means known then; hollow means not yet observed. A retrospective row can hold a future label without making that label available for the earlier decision.

A model may legitimately learn from completed historical pairs of inputs and forward returns. For a nonzero starting price, the label

rt,t+h=Pt+hPt−1r_{t,t+h}=\frac{P_{t+h}}{P_t}-1

is aligned to the starting row so it can be matched to inputs available there. That alignment is useful for supervised learning; it does not permit using rt,t+hr_{t,t+h} as one of those same inputs. Nor does a price-return label establish that a strategy could transact at PtP_t after observing it. Feasible entry timing is a separate part of the trading rule.

The two clocks to check

For each prediction row, distinguish the decision time from the label-completion time.
  • A predictor must be available by the decision time. A historical observation date is insufficient if the value was published, received or revised later.
  • A target must be available by the model-fit time before that row can enter the training set. At a refit at t=0t=0, the row-00 forward return above is still unfinished.

The same row can therefore be eligible for prediction now and eligible for training only later. Sorting a dataset chronologically does not, by itself, enforce either condition.

Observation, Event and Availability Timestamps develops the first clock, including publication delays and same-bar execution errors. Labels and Overlapping Outcomes develops the second, including unfinished labels and shared return intervals. Both have public previews; full lessons require membership.

Future-dated does not always mean unknown

A future calendar value can already be known. For example, the day of the week at the forecast horizon can be supplied to a model today. A scheduled event indicator can also be usable if its schedule was actually known then; later schedule changes must not be backfilled into the earlier information set.

By contrast, a future realised price or the eventual numerical surprise in an economic release is not known merely because its observation date appears in the dataset. An advance forecast of that quantity is a different variable, with its own issue time and forecast error. Forecasting with realised future predictors is an ex-post exercise; it does not demonstrate an ex-ante trading forecast.

A lead is not a prediction—or a leading signal

The realised lead yt+hy_{t+h} and a forecast y^t+h∣t\hat y_{t+h\mid t} refer to the same future observation but play different roles. The first is the outcome; the second is an estimate made using information available at tt. Their difference is a forecast error once the outcome arrives.

Likewise, saying that a market signal “leads returns” claims a predictive relationship: today's signal is associated with later returns. Constructing a lead column establishes no such relationship and no causal effect. A lagged variable, yt−hy_{t-h}, moves in the opposite direction and refers to an earlier observation; its actual availability still needs checking.

In implementation, define the clock before shifting. One step may mean the next trading session, the next event or a fixed calendar interval. Sort within each instrument, respect missing observations, and verify which timestamps remain attached to the shifted values. A next-row lead in an irregular table is not automatically a one-day lead.

Sources

  • Chris Brooks, Introductory Econometrics for Finance, second edition, Cambridge University Press, 2008, §§5.11.3–5.11.4, pp. 245–248: forecast horizons, information available at the forecast origin and unknown future regressors.
  • Rob J. Hyndman and George Athanasopoulos, Forecasting: Principles and Practice, third edition, §7.6: ex-ante and ex-post forecasting, known calendar predictors and horizon-aligned predictive regression.
Last updated
  • time series
  • forecasting
  • target alignment
  • look ahead bias
Keep connecting the ideas
← Back to the glossaryBack to top ↑