Time series regression: trends, drivers and checks

Time series regression relates a dated outcome, such as monthly revenue, to trend and business drivers while respecting the order of observations. It can help finance explain changes and assess supplied predictions, but a strong historical fit is not proof of a useful forecast. Check timing, seasonal patterns, residual dependence and performance on later periods before acting on the result.

What the relationship measures

A simple specification separates a baseline, a time trend and a business driver:

Revenue in period t = Baseline + Trend coefficient × t + Driver coefficient × Driver in period t + Error in period t

The time index marks successive periods. The driver might be contracted deliveries, active accounts or another measurable operating input. A coefficient describes an estimated relationship while holding the included predictors fixed; it does not prove that changing the driver will cause the estimated revenue change.

Time order adds structure. The NIST introduction to time series analysis identifies autocorrelation, trend and seasonal variation as features that need attention. A revenue series can rise with a driver simply because both grow over time. An analyst needs to examine that possibility before interpreting their relationship as business evidence.

For the wider process of inspecting dated observations, see time series analysis. Use time series data correlation for the separate question of whether dated series move together. The focus here is the regression relationship and the checks around its supplied results.

Build a review file with dates that mean the same thing

Use one row per reporting period, with the actual outcome, driver values, supplied fitted or forecast values, and an indicator showing which observations were used for estimation. Keep the forecast issue date and horizon where applicable. They establish what information was available at the time.

Reconcile actual revenue to the approved ledger export. Confirm that all drivers use the same monthly cutoff and entity scope. A missing month needs investigation; deleting its row can turn a lag into the wrong calendar relationship.

Keep historical fitted values separate from forecasts. A residual is actual minus fitted outcome within the estimation data. A forecast error is actual minus a prediction for an observation withheld from fitting. Mixing these into a single accuracy total hides whether the specification has faced unseen periods.

The regression worksheet guide covers implementation mechanics. This review starts with supplied coefficients and predictions, so the business team can check their meaning and reconcile the results.

A worked revenue forecast review

Consider a fictional US service company. Before January 2026, its planning analyst supplies coefficients estimated using earlier history: a baseline of $80,000, a monthly trend contribution of $5,000 and $2,000 per scheduled contract. The coefficients are illustrative inputs, not estimates fitted to the six rows below.

The analyst freezes this forecast on December 31, 2025. Contract counts are the signed contracts scheduled for delivery in each future month as known on that date. Later cancellations or newly signed business do not overwrite those frozen driver inputs.

Supplied monthly prediction (USD) = 80,000 + 5,000 × t + 2,000 × Scheduled contracts

The prediction file reconstructs as follows:

Compare the frozen predictions with the later approved actuals:

The reconciliation is $839,000 − $869,000 = −$30,000, equal to the sum of monthly errors. Mean absolute error is $30,000 / 6 = $5,000 per month. These calculations use whole dollars without rounding.

Every forecast error is negative, and the misses generally widen over time. That supports investigating persistent overprediction. It does not, by itself, establish statistical autocorrelation or identify the cause. Check contract cancellations, delivery timing, pricing and the assumption of a continuing linear trend.

Check trend, seasonality and lagged effects

Trend. A fixed monthly increase can help summarize history but become implausible when capacity, prices or demand change. Ask whether the trend is standing in for an omitted operating driver. The example includes trend explicitly so its contribution is visible.

Seasonality. If the same calendar months repeatedly behave differently, a trend alone will not describe that pattern. An analyst can include seasonal terms, such as month indicators with a reference month. Inspect multiple annual cycles before drawing a seasonal conclusion; the six review months above cannot establish a full annual pattern.

Lags. A lead created this month may become revenue later. A lagged driver links the outcome to an earlier driver's value. Choose the lag using business timing and chronological testing, then document it. Future realized drivers are unavailable at an earlier forecast date unless they were already known or explicitly supplied as assumptions.

Residual dependence. On the estimation history, inspect residuals in date order and their autocorrelation. Repeated patterns suggest the specification leaves time structure unexplained. Forecasting: Principles and Practice on regression discusses residual checks and regression predictors. Ask the analyst to review the specification and uncertainty before using its intervals for commitments.

Validate in chronological order

Estimate on earlier observations, freeze the specification, and evaluate on later periods. For repeated checks, move the forecast origin forward and keep the training history before each test observation. This is the rolling-origin approach described by the forecasting text's authors.

Match the test to the decision. The worked example measures a forecast issued once before January, so later months have longer horizons. Do not label its results as a series of freshly updated monthly forecasts. Compare alternative specifications and simple benchmarks using the same issue dates, information and horizons.

FP&A owns the reconciliation and the planning decision; the analyst responsible for estimation owns the statistical specification. The broader time series forecasting guide covers the forecast workflow, while ARIMA time series analysis addresses a different way to handle time dependence.

Where it goes wrong

• Randomly mixed training and test months. Later observations can inform what is presented as an earlier forecast. Preserve the chronology and the information cutoff.
• Revised drivers replace frozen inputs. A forecast reviewed against information learned later no longer represents the original decision. Retain the issued version.
• Shared trend becomes a causal story. A rising driver and rising revenue can move together without the driver causing the outcome. Review omitted variables and timing.
• Residual patterns are ignored. A small historical error total can coexist with persistent runs or seasonal misses. Inspect errors in date order.
• Actuals and predictions use different scopes. Missing entities, currency changes or revised revenue cutoffs create apparent forecast misses. Reconcile those differences first.

Review supplied regression results in Covirage

Upload your dated revenue, driver and supplied prediction files for Covirage to map the columns and its tools to compute and check the comparisons. The external AI model explains the results with every figure traced to its rows. Create a free Covirage account and bring a sample of your own forecast review file; creating the account takes about a minute.

See how this supports forecast analysis.