A lagged variable is the value of a variable from an earlier observation period, aligned with a later period being analysed. In a time series, is the one-period lag of , and is its -period lag for a positive integer . In quant research, that alignment determines whether a forecast uses history that was actually available when the decision was made.
A lag retains the original variable's units. Yesterday's price is a lagged price; it is not yesterday's return or the change since yesterday. A first difference subtracts the lagged level from the current level.
Read the row from the target's point of view
Take three synthetic index levels, observed at equally spaced times:
| Target time | Current level | One-period lag | Two-period lag |
|---|---|---|---|
| 0 | 10 | Unavailable | Unavailable |
| 1 | 12 | 10 | Unavailable |
| 2 | 15 | 12 | 10 |
The value 12 appears in the current-level column at time 1 and in the one-period-lag column at time 2. The observation has not changed; its alignment has. A lag operator expresses this compactly as .
The initial missing entries are unavailable history, not zero. Starting with consecutive observations and requiring lags through , a complete-case dataset has usable rows when , assuming no other missing values or earlier history. Grouped asset data require lags within each asset; the previous row of a combined table might belong to another security.
A forecast has two times
Suppose each level becomes usable immediately at its observation time. Immediately after observing , issue a forecast for time 2 using the stipulated rule
The subscript means target time 2, information available at time 1. Relative to the target, the input is the one-period lag. Relative to the decision, it is the latest observation. The intercept is 1 index point and the slope 0.5 is dimensionless. These coefficients are toy assumptions, not estimates or evidence of a successful trading model.
When time 2 arrives, . With the error defined as actual minus forecast,
The lagged input is 12, the forecast is 7, and the later outcome is 15. Substituting 15 into the predictor at the time-1 decision would use the future target. The actual change, index points, is a fourth object: the first difference.
Original synthetic index-level example. Observations are equally spaced and usable immediately at their observation time. Filled markers are known at the time-1 decision; the hollow target is revealed later. All values and errors are in index points; coefficients are stipulated.
Earlier in the table can still mean unavailable
A period label describes what the observation measures, not necessarily when it could be used. A quarterly figure can refer to a completed quarter but be released after the researcher's next decision. A later revision can also change the value stored against an old period. ALFRED's vintage-data framework explicitly distinguishes past observation periods from the dates on which new or revised data became available.
Consequently, shifting a final revised macro series by one row does not automatically produce a historically feasible predictor. Use the vintage and release or receipt time that the strategy could access, and allow for processing before the decision. Otherwise a seemingly lagged feature can contain look-ahead bias.
The same issue arises with irregular records: a one-row lag means the previous recorded observation, which may be several sessions earlier. State the intended clock—calendar day, trading session or event—and handle missing periods explicitly. Observation, Event and Availability Timestamps develops these timing checks (public preview; full lesson requires membership).
What including a lag does—and does not—say
An autoregressive model uses the outcome's own earlier values as predictors. A distributed-lag specification uses earlier values of another variable, for example allowing previous policy-rate changes to help explain a current response. A lag is therefore an input alignment, not a claim that every time series follows an autoregression.
Nor does “earlier” establish causation or remove correlation between a regressor and the model's disturbance. The choice and number of lags still require economic justification, appropriate estimation assumptions and evaluation on later data. AR, MA, and ARMA connects lag notation to forecast models and diagnostics (public preview; full lesson requires membership).
Finally, keep the forecast origin fixed when evaluating a multi-step forecast. An intermediate realised observation may be used for a new forecast after it becomes available; it cannot be silently inserted into the forecast issued earlier. Rolling one-step forecasts and a fixed-origin multi-step path answer different questions.
Sources
- Chris Brooks, Introductory Econometrics for Finance, second edition, Cambridge University Press, 2008, §4.5.1, p. 140; §5.12, pp. 256–257: lag construction, missing initial observations and forecast origins.
- Rob J. Hyndman and George Athanasopoulos, Forecasting: Principles and Practice, third edition, §§9.2–9.3: backshift notation and autoregressive predictors; §10.6: delayed effects of other predictors.
- Federal Reserve Bank of St. Louis, ALFRED Help: vintage observations, revisions and release dates. The numerical example and timeline above are original synthetic constructions.