Forecasting is a means, not a trophy
A forecast exists to make a decision better. If it does not change what anyone does, it is decoration. Over five sessions you have learned what a series is, the ladder of methods, who builds it, and how to measure it. This capstone turns that into leadership judgement: choose the right horizon, avoid the traps that make good-looking models fail, recognise the series you should leave alone, and grow capability at a sustainable pace. The through-line is honesty - a confident forecast of an unforecastable thing is worse than no forecast, because someone will act on it.
Match the horizon to the decision 5 min live
The first strategic choice is not which model - it is how far ahead and how often. A staffing decision needs tomorrow at daily resolution; a capacity investment needs next year at monthly. The rule that never changes: the further out you forecast, the more uncertain it gets. Push the horizon past what the data can honestly support and you are not planning, you are guessing with extra steps.
LiveHorizon, frequency, and the cost of both errors3 min▶
- Start from the decision, work back. What choice does this forecast inform, and when is it made? That fixes the horizon and frequency - not the other way round.
- Longer horizon = wider interval. Next week is fairly knowable; next year much less. Accept it, and plan with the range, not the point.
- Both directions of error cost money. Forecast too high and you over-stock, over-staff, over-build. Too low and you stock out, miss demand, disappoint. Ask which error is more expensive here - it changes the whole design.
Lumen wants "a forecast". The useful question is which one. Daily sessions 14 days out drives ad and support staffing. Monthly signups a year out drives the hiring plan. Same brand, two different forecasts, two horizons, two acceptable error bars. Bundling them into one "the forecast" is how projects quietly fail.
The four classic pitfalls 7 min live
Most forecasting failures are not exotic. They are the same handful of traps, over and over. You do not need to catch them in the code - you need to recognise the symptom and ask the question that exposes each one.
Self-studyRegime change - when the past stops predicting the future2 min read▶
The deepest pitfall is not in the model at all. Every forecast assumes tomorrow resembles yesterday. When the world shifts - a new pricing model, a competitor's launch, a pandemic, a policy change - that assumption breaks, and history becomes a poor guide to what comes next. This is a regime change.
- No model rescues you from a regime change. A more complex model fit to the old regime just fails more confidently.
- The interval is the early warning. When live errors keep landing outside the prediction interval, the regime may have shifted - time to re-examine, not to re-tune.
- Judgement re-enters here. After a genuine break, human context about why often beats any automated forecast until enough new data accrues.
Self-studyCold-start - forecasting a series with no history2 min read▶
A brand-new product, store, or campaign has no past to learn from. You cannot fit a model to data that does not exist - but you are not helpless.
- Analogs. Borrow the shape from a similar existing series (a comparable product's first 90 days) and adjust.
- Hierarchy. Forecast the group you do have history for (the category), then allocate a share to the new item.
- Judgement, stated as a range. An honest expert estimate with a wide interval beats a precise-looking model trained on nothing.
When NOT to forecast 6 min live
The most senior thing a leader can say in a forecasting review is "we should not forecast this". It protects the team from a false number that someone will inevitably act on. There are three honest reasons to decline.
LiveToo little data, a regime change, or a rule is enough3 min▶
- Too little data. A handful of observations cannot reveal a trend or a seasonal pattern. Any "forecast" is really an opinion with a chart. Wait, use analogs, or decide by judgement.
- A regime change. If the process that generated the past is gone, the past cannot predict the future. Forecasting across a genuine break manufactures false confidence.
- A simple rule is good enough. If "order the same as last month" or a fixed safety buffer already makes the decision well, a forecasting project adds cost and fragility for no decision gain. The best forecast is sometimes no forecast.
Honesty beats a confident guess. Presented with a brand-new SKU and asked for a 12-month demand forecast, the wise answer is: "We have two weeks of data through a launch spike - any monthly number would be fiction. Here is an analog-based range, we will replace it with a real forecast once we have a clean season." A confident forecast of an unforecastable series is worse than none, because it looks trustworthy.
Self-studyBuild vs buy - a brief leader call2 min read▶
Once forecasting is worth doing, should you build in-house or buy a managed service? A quick frame:
- Buy when forecasting is not your differentiator, the series are standard, and you want speed. A managed service handles the plumbing so a small team can ship.
- Build when forecasting is core to the business, you need custom logic or tight integration, and you have the team to own it long term.
- Either way, you still own the honesty. Backtesting, beating naive, and reporting intervals are your responsibility whether the model is yours or a vendor's. A bought forecast is not exempt from the four checks.
A crawl-walk-run roadmap 5 min live
You do not need deep models on day one. Forecasting capability grows in stages, and each stage must earn the next by proving value on a backtest. Start with the baseline, add smoothing and Prophet, make it honest with backtests and intervals, reach for ML only where it pays, then close the loop with monitoring.
Try it yourself - this week ◐ 25-35 min total
- Draft your org's four-step forecasting roadmap: name the baseline you would start from, the first real method, how you would backtest and add intervals, and what "run" looks like (ML or a monitored loop).
- For each step, write the one decision it unlocks - if a step unlocks no decision, cut it.
- Name one series your org should NOT try to forecast, and write why (too little data, a recent regime change, or a simple rule already suffices).
- Pick one live forecast and run it against the four pitfalls: any leakage risk, any over-fitting, is there an interval, is anyone chasing accuracy past its value?
- Optional: skim FPP3 (Hyndman & Athanasopoulos, otexts.com/fpp3) on judgemental forecasting and the M-competition findings on simple methods.
Sources covered (cited)
Taught from the forecasting canon and the published empirical evidence. This session covers:
Three questions before you go 🎯 ◐ 90 seconds
1 · What should set the horizon and frequency of a forecast?
Start from the decision and work back. A longer horizon means a wider interval; push past what the data supports and you are guessing, not planning.
2 · Which is an honest reason NOT to forecast a series?
A confident forecast of an unforecastable series is worse than none, because someone will act on it. Sometimes the best forecast is no forecast.
3 · In the crawl-walk-run roadmap, what must never be skipped?
Each step earns the next on a backtest. A model that cannot beat naive on hidden data has not proven its complexity and should not advance.