Forecasting

Forecasting estimates future values from historical data, current information, assumptions, models, and judgment, with explicit uncertainty and error review.

Forecasting is the disciplined process of estimating future values from historical data, current information, assumptions, models, and judgment. In finance and business, common targets include sales, cash receipts, costs, losses, prices, rates, demand, and economic variables.

A forecast is an estimate, not a target or promise. Its value comes from supporting a decision, quantifying uncertainty, and improving as errors are reviewed.

Key Takeaways

  • Define the variable, unit, frequency, horizon, and decision before choosing a method.
  • Separate the forecast from budgets, targets, scenarios, and commitments.
  • Use only information that would have been available at the forecast date when backtesting.
  • Compare complex methods with simple benchmarks.
  • Report uncertainty, assumptions, and known break risks.
  • Evaluate bias and decision impact, not only average statistical accuracy.

Forecast, Budget, Target, and Scenario

ItemMain questionShould it be the most likely outcome?
ForecastWhat is expected under stated information and assumptions?Usually a central estimate or distribution
BudgetWhat resources and spending are authorized or planned?Not necessarily
TargetWhat result does management seek or require?No; it can be deliberately challenging
ScenarioWhat happens under a coherent alternative state?Not necessarily
CommitmentWhat amount or action is contractually promised?No; it is an obligation, not an estimate

Using a sales target as the cash forecast can overstate liquidity. Using a conservative budget as the unbiased forecast can distort performance evaluation.

Forecasting Workflow

  1. Define the decision. Specify what action the forecast informs and the cost of being wrong.
  2. Define the target. Record unit, timing, population, currency, and accounting basis.
  3. Freeze the information set. Preserve source files, revisions, and assumptions available on the forecast date.
  4. Explore the data. Check missing values, outliers, trend, Seasonality, structural breaks, and changing definitions.
  5. Choose benchmarks. Include naive, same-period-last-year, moving-average, or simple driver models.
  6. Estimate candidate models. Keep model complexity proportional to data and decision value.
  7. Validate out of sample. Use rolling or expanding windows without future information leakage.
  8. Add documented judgment. Record overrides separately from model output.
  9. Publish uncertainty. Provide ranges, scenarios, or distributions appropriate to the use.
  10. Review outcomes. Measure errors, diagnose causes, and update the process.

Common Methods

MethodUseful whenMain risk
Naive or seasonal naiveA stable persistence or seasonal benchmark is neededMisses trend changes and turning points
Moving AverageShort-run noise should be smoothedLags abrupt changes
Exponential smoothingLevel, trend, and seasonality evolve graduallyPerforms poorly after structural breaks
Regression or driver modelCredible business or economic drivers are availableCorrelation may not remain stable or causal
ARIMA or other time-series modelAutocorrelation and lag structure contain informationSensitive to specification and regime change
Machine learningRich nonlinear predictors and enough observations existLeakage, overfitting, and weak interpretability
Judgmental forecastNovel events are absent from historical dataBias and inconsistency unless overrides are logged

Combining forecasts can reduce dependence on one model, but it does not remove shared data or assumption risk.

Worked Example: Sales to Cash Receipts

Assume current quarterly sales are $10 million. The forecast uses:

  • unit-volume growth of 3%;
  • average-price growth of 2%; and
  • 92% collection within the quarter.

Forecast sales are:

$$ \$10.0\text{m}\times1.03\times1.02=\$10.506\text{m} $$

Expected same-quarter cash receipts from those sales are:

$$ \$10.506\text{m}\times92\%=\$9.6655\text{m} $$

The sales forecast and cash forecast are different targets. Existing receivables, taxes, returns, bad debts, and timing within the quarter may also affect cash. If the decision is a covenant or payroll test, the model should forecast the relevant cash dates rather than stop at revenue.

Forecast Error Measures

Let the error be actual minus forecast:

$$ e_t=y_t-\hat{y}_t $$

Useful summaries include:

$$ \text{Mean error}=\frac{1}{n}\sum_{t=1}^{n}e_t $$
$$ \text{MAE}=\frac{1}{n}\sum_{t=1}^{n}|e_t| $$
$$ \text{RMSE}=\sqrt{\frac{1}{n}\sum_{t=1}^{n}e_t^2} $$

Mean error reveals directional bias. MAE is easy to interpret in the target’s units. RMSE penalizes large misses. Percentage errors can fail when actual values are zero, near zero, or negative.

The best metric depends on the decision. A liquidity forecast may penalize overforecasting cash more heavily than underforecasting it.

Backtesting Without Leakage

A valid historical test should mimic real forecasting:

  • train only on data available before each forecast date;
  • preserve original data vintages when revisions matter;
  • estimate seasonal factors within each historical window;
  • keep model-selection and parameter-tuning data separate from final testing;
  • include failed periods and regime changes; and
  • compare all models over the same origins and horizons.

Randomly splitting a time series can leak future patterns into the training set. Time-ordered rolling or expanding windows are usually more appropriate.

Forecast Uncertainty

Uncertainty comes from future shocks, parameter estimates, model choice, data error, and conditioning assumptions. Communication options include:

  • prediction intervals with a defined probability interpretation;
  • judgmental ranges that clearly lack one;
  • discrete upside and downside scenarios;
  • sensitivity tables for key drivers; and
  • reverse stress tests identifying the assumption that breaks a limit.

A narrow model interval is not complete if it excludes model and scenario uncertainty.

Governance and Review

Good forecast records preserve:

  • forecast date and preparer;
  • target definition and horizon;
  • source-data vintage;
  • code or model version;
  • baseline assumptions;
  • model output and manual overrides;
  • approved final forecast and range;
  • realized outcome definition; and
  • error diagnosis and corrective action.

This audit trail prevents teams from silently replacing an old forecast with a revised one after the outcome becomes known.

Main Risks and Limitations

  • Structural breaks: relationships change after policy, technology, competition, or crisis shifts.
  • Data quality: missing, revised, or inconsistent observations weaken estimates.
  • Overfitting: a model can explain noise in historical data.
  • Leakage: future information can contaminate training or validation.
  • Bias: targets, incentives, and group pressure can influence judgment.
  • Aggregation: a strong total forecast can hide severe segment errors.
  • Tail risk: ordinary error bands may omit rare but material events.
  • Decision mismatch: the statistically best forecast may not minimize financial cost.

Common Mistakes

  • Saying forecasting only uses historical data.
  • Confusing the forecast with the budget or target.
  • Selecting the model on in-sample fit alone.
  • Ignoring a simple seasonal-naive benchmark.
  • Revising old forecasts after outcomes arrive.
  • Reporting percentage accuracy for near-zero values.
  • Forecasting revenue when the decision depends on cash or liquidity.

Authoritative Sources

FAQs

What is the simplest useful forecast benchmark?

It depends on the series. Common benchmarks are the last observed value, the same period one year earlier, or a simple trend. A complex model should demonstrate improvement over a relevant simple benchmark.

Is the most accurate forecast always the best decision forecast?

No. The costs of over- and underforecasting can differ, and a slightly less accurate model may be more stable, timely, explainable, or useful for the decision.

How often should a forecast be updated?

Update frequency should match new information and the decision horizon. Reforecasting constantly on noise can be as harmful as ignoring material changes.

This page is educational and does not provide accounting, economic forecasting, investment, credit, or business advice.

Browse Economics