📘 How does forecasting turn history into a guess?

What you’ll learn

1. What Forecasting Is — and What It Cannot BeDefine forecasting, distinguish it from prediction and planning, and frame what makes a series forecastable.Forecasting is the disciplined estimation of future values of a variable using current and historical information, and it differs from a target (what you want) or a plan (what you will do). Forecastability depends on understanding the drivers, having relevant data, and whether the act of forecasting itself changes the outcome. A central principle is that every forecast is uncertain, so a useful forecast quantifies that uncertainty with a prediction interval rather than offering a single number.
2. Reading a Time Series: Patterns, Decomposition, and the Case BeginsIdentify trend, seasonality, and cyclic patterns in a time series and decompose a series into interpretable components.Before modeling, analysts must understand the structure of the data through visualization and decomposition. A time series can carry trend (long-run direction), seasonality (a fixed, calendar-driven pattern of known period), and cyclic behavior (longer fluctuations of no fixed length, such as business cycles), plus irregular remainder. Decomposition separates these components so each can be understood and, if needed, modeled — and it is the foundation for the running case study of a fictional coffee retailer, BrewMetric Co.
3. Qualitative and Judgmental ForecastingApply structured judgmental forecasting methods appropriately and recognize the cognitive biases they are designed to control.When little or no relevant historical data exists — new products, new markets, or disruptive events — judgmental methods become necessary, but unstructured judgment is vulnerable to systematic bias. Structured approaches such as the Delphi method, scenario forecasting, and forecasting by analogy impose discipline that reduces anchoring, optimism, and groupthink. Even in data-rich settings, judgment is used to adjust statistical forecasts for information the model cannot see, and these adjustments must themselves be documented and reviewed.
4. Quantitative Methods I: Benchmarks, Smoothing, and RegressionApply baseline benchmark methods, the exponential smoothing family, and explanatory regression to produce quantitative forecasts.Quantitative forecasting splits into time-series methods, which extrapolate a variable's own history, and explanatory methods, which relate the target to predictor variables. Simple benchmarks — naive, seasonal naive, mean, and drift — are essential because no sophisticated method should be adopted unless it beats them. Exponential smoothing methods build forecasts from weighted averages of past observations, extending to capture trend and seasonality, while regression connects demand to drivers such as price and promotions for both forecasting and what-if analysis.
5. Quantitative Methods II: ARIMA, Method Selection, and CombinationExplain the ARIMA family, choose between method classes for a given series, and justify forecast combination.ARIMA models describe a series through its autocorrelations, combining autoregression, differencing for stationarity, and moving-average terms, with seasonal extensions for calendar effects. ETS and ARIMA are complementary frameworks that overlap but are not identical, and choosing between them is an empirical question answered by out-of-sample testing rather than dogma. A robust, evidence-backed strategy is to combine forecasts from several methods, which large-scale competitions have repeatedly shown tends to outperform individual models.
6. Evaluating Accuracy and Building Your Forecast ArtifactEvaluate forecasts with appropriate error metrics and time-series validation, then build a mini forecasting artifact for BrewMetric and defend it.A forecast is only trustworthy if it is evaluated honestly on data the model has not seen, using time-series cross-validation rather than a random split that would leak future information. Accuracy metrics each answer a different question — MAE and RMSE are scale-dependent, MAPE is a percentage but fails near zero, and scaled errors like MASE allow comparison across series — while bias and the tracking signal detect systematic over- or under-forecasting. In the capstone you assemble these pieces into a small, defensible forecasting artifact and the simulation scores how well your method, validation, and chosen metrics fit BrewMetric's decision.

Questions this course answers

Grounded in trusted sources

• Hyndman, R.J., & Athanasopoulos, G. (2021). Forecasting: Principles and Practice (3rd ed.), Ch. 1. OTexts. https://otexts.com/fpp3/
• Makridakis, S., Spiliotis, E., & Assimakopoulos, V. (2018). The M4 Competition: Results, findings, conclusion and way forward. International Journal of Forecasting, 34(4), 802-808.
• Hyndman, R.J., & Athanasopoulos, G. (2021). Forecasting: Principles and Practice (3rd ed.), Ch. 2-3 (Time series graphics; Time series decomposition). OTexts. https://otexts.com/fpp3/
• Hyndman, R.J., & Athanasopoulos, G. (2021). Forecasting: Principles and Practice (3rd ed.), Ch. 6 (Judgmental forecasts). OTexts. https://otexts.com/fpp3/
• Hyndman, R.J., & Athanasopoulos, G. (2021). Forecasting: Principles and Practice (3rd ed.), Ch. 5, 7-8 (The forecaster's toolbox; Time series regression models; Exponential smoothing). OTexts. https://otexts.com/fpp3/
• Hyndman, R.J., & Athanasopoulos, G. (2021). Forecasting: Principles and Practice (2nd ed.), §7.3 Holt-Winters' seasonal method. OTexts. https://otexts.com/fpp2/holt-winters.html

Every Wunder lesson is built from real, reputable sources — never invented.

Related courses