Forecasting
Also known as: demand forecasting, revenue forecasting, sales forecasting, predictive forecasting, forecast accuracy, MAPE
Forecasting is the process of estimating future values of a variable based on historical data, statistical models, and/or expert judgment.
Business forecasting encompasses revenue forecasting (what will we earn next quarter?), demand forecasting (how many units will customers want?), workforce planning (how many staff will we need?), and economic forecasting (what will macroeconomic conditions look like?). Forecasts are always uncertain -- the goal is not perfect prediction but quantifying the range of plausible outcomes to support better planning decisions. A good forecast includes not just a point estimate (single predicted value) but a confidence interval representing the range of likely outcomes.
Forecasting, explained at three levels
At home
Forecasting is making an educated guess about what will happen in the future, based on what has happened in the past and any other relevant information. Weather forecasting uses temperature, pressure, and historical weather patterns to predict tomorrow's conditions. Business forecasting uses last year's sales, growth rates, and seasonal patterns to estimate next quarter's revenue. No forecast is guaranteed to be right -- but a good forecast gives you a reasonable range of expectations so you can plan appropriately, whether that means stocking enough inventory, hiring enough staff, or setting a realistic budget.
At work
Business forecasting produces the numbers that drive planning: revenue forecasts determine budget allocations; demand forecasts determine inventory orders; staff forecasts determine hiring plans. A company that forecasts 20% revenue growth will plan very differently from one forecasting flat growth. The value of forecasting is not in the specific numbers (which are always wrong to some degree) but in the planning discipline: forcing explicit assumptions about the future, creating shared expectations across departments, and enabling early detection when actual results diverge from plan (allowing course correction).
At home
Forecast accuracy metrics: Mean Absolute Error (MAE -- average of absolute differences between forecast and actual); Mean Absolute Percentage Error (MAPE -- average of percentage differences, making it comparable across different scales); Root Mean Square Error (RMSE -- penalizes large errors more heavily). MAPE is the most intuitive for business communication ('our forecasts are off by an average of 12%') but has problems when actual values can be zero or close to zero (division by zero; very sensitive to small actual values). RMSE is more appropriate when large errors have disproportionately bad consequences.
At work
Forecast uncertainty and confidence intervals: a point forecast (single number) is less useful than a forecast with a confidence interval showing the range of plausible outcomes. A revenue forecast of $10M +/- $2M (80% confidence interval: $8M-$12M) tells planners to stress-test plans at both $8M and $12M scenarios. Forecast uncertainty grows with the forecast horizon (forecasting next week is more accurate than forecasting next year), with volatility in the underlying metric, and with the quality of available predictor variables. Presenting forecasts as ranges rather than precise point estimates is more honest and more useful for decision-making.
At home
Forecast bias vs. forecast accuracy: a forecast can be unbiased (correct on average across many forecasts) but highly inaccurate (individual forecasts vary widely from actuals). Or it can be biased (consistently too high or too low) but precise (individual forecasts cluster tightly around the biased average). In business forecasting, bias is often worse than noise: a systematically optimistic revenue forecast leads to systematic over-staffing and over-spending; a systematically pessimistic demand forecast leads to systematic stockouts and lost sales. Tracking forecast bias separately from accuracy (using metrics like Mean Forecast Error, not just absolute error) identifies systematic over- or under-forecasting tendencies that accuracy-only metrics miss.
At work
Judgmental forecasting and its integration with statistical models: pure statistical models miss information that forecasters have about planned events (a major marketing campaign, a competitor's product launch, supply disruption, regulatory change). Pure judgmental forecasts miss statistical patterns and are prone to human biases (optimism, anchoring on recent events, political pressure to forecast favorably). Best-in-class forecasting processes combine both: use statistical models as a baseline, then allow structured judgmental adjustments for planned events and known contextual factors not captured in historical data. Structured adjustment processes (requiring explicit documentation of adjustment rationale) improve accountability and enable learning from past adjustments.
Real-world examples
At home
Lena was planning her freelance income for the year. She looked at her client history over the past three years and found: income was consistently higher in Q2 and Q4 (when clients had budget remaining and were spending before year-end) and lower in Q1 (January-March was typically slow). She used this pattern to create a monthly income forecast, planning to save more in high-income months to cover low-income months, and to reduce expenses in Q1. Her forecast wasn't perfect but it gave her a realistic budget plan that matched her actual cash flow pattern.
At work
Thomas led supply chain for a consumer goods company. His team built a demand forecasting model using: historical sales data (12 months), promotional calendar (planned discounts and campaigns), seasonality factors (estimated from 3 years of data), and economic indicators (consumer confidence, competitive category trends). The model produced weekly item-level forecasts for 500 SKUs with 90% confidence intervals. Inventory orders placed 8 weeks in advance used the forecast plus safety stock sized to the confidence interval width. Service levels improved significantly from ad-hoc ordering based on recent sales alone.
Common confusions
- Many organizations treat forecasts as commitments rather than probability estimates — A forecast is a probabilistic estimate of the future -- it is correct that it will be wrong to some degree, and treating it as a commitment creates harmful incentives (sandbagging to ensure the 'commit' is achievable, political pressure to forecast favorably, resistance to updating forecasts as new information arrives). A better culture treats forecasts as calibrated probability estimates: teams are evaluated on forecast accuracy (how often did actuals fall within forecast confidence intervals?) rather than on whether they hit a point forecast. This creates incentives for honest forecasting rather than conservative or optimistic positioning.
- Some assume that more complex forecasting models always produce more accurate forecasts — The M-competitions (empirical forecasting accuracy comparisons across many methods) consistently found that simple methods (exponential smoothing, seasonal naive) outperform or match complex statistical and ML methods on many business time series. Complex models overfit to historical data and perform worse on new data; they require more data to train and more expertise to validate; and their black-box nature makes it harder to incorporate judgmental adjustments. The practical recommendation is to start with simple benchmarks (seasonal naive, exponential smoothing) and only adopt more complex methods if they demonstrably outperform on held-out test data.
For analysts
Forecasting is a high-value service for clients in planning-intensive industries: retail (demand and revenue forecasting), financial services (revenue and risk forecasting), manufacturing (demand and capacity planning), healthcare (patient volume and staffing forecasting), and any organization with significant inventory, staffing, or capital expenditure decisions. The practical entry point is a forecast accuracy audit: how accurate have the client's current forecasts been? Measuring historical forecast bias (consistently too high or low?) and accuracy (MAPE) establishes the baseline and identifies whether better forecasting methodology, better input data, or better process (structured judgment) would add the most value.
History & origin
Forecasting has been practiced in agriculture, trade, and military planning throughout human history. Modern statistical forecasting developed alongside statistical methodology in the 20th century. Exponential smoothing methods (Holt, 1957; Holt-Winters, 1960) provided computationally simple but effective forecasting for business time series. Box-Jenkins ARIMA methods (1970) provided a more comprehensive statistical framework. The M-competitions (Makridakis Competitions, starting 1982) established empirical benchmarks for forecasting accuracy across methods -- finding that simple methods often outperformed complex statistical models in practice. The M4 (2018) and M5 (2020) competitions included machine learning methods at scale, finding hybrid approaches (combining statistical and ML methods) performed best. Meta's Prophet (2017) made production-quality forecasting accessible to non-statisticians.