Time-series fundamentals

Before AI can help you forecast, you need the vocabulary forecasts are built from — because the most common forecasting failure isn't a bad model, it's a fuzzy question. Every business time series decomposes into four parts, and saying which part you're arguing about instantly upgrades any forecast meeting:

• Level — where the series sits now. Alder's weekly freight revenue: ~$2.1M. The anchor everything else modifies.
• Trend — the persistent drift: up 4% year-over-year as two new contracts ramp. Trends deserve explanations, not just extrapolation — a trend you can't attribute to something real (contracts, pricing, capacity) is a pattern waiting to betray you.
• Seasonality — the repeating shape: freight peaks pre-holiday (weeks 38-48), sags in January; AP disbursements spike at month-end; payroll lands biweekly like clockwork. Seasonality is the most learnable component — a year of history usually reveals it — and ignoring it is the most embarrassing forecast error because everyone in operations already knew.
• Events & noise — the residue: the week a customer's plant shut down, the diesel spike, the random wobble. The discipline is separating explainable events (annotate them! this is what the forecast log is for) from noise (accept it — chasing noise with model complexity is how forecasts get worse while looking smarter).

Naive baselines: the honesty floor

Before any clever method, compute the dumb ones: last period's value (naive), same period last year (seasonal naive), and trailing 4-week average. These cost nothing and set the bar: any forecasting effort — human, statistical, or AI — justifies itself only by beating the naive baseline consistently, measured on held-out history. You'd be surprised (or, if you've worked in FP&A, not surprised at all) how often an elaborate forecast loses to 'same week last year, plus trend.' The baseline isn't an insult to the analyst; it's the control group that makes 'our forecast works' a measurable claim instead of a vibe. Keep a running score: forecast vs. actual vs. naive, every period, in one small table.

Grain and horizon: decide before you model

• Grain — forecasting weekly cash is a different problem than monthly revenue than daily volume by lane. Finer grain = more noise, more rows, more operational usefulness; coarser = smoother, strategic, hides timing crunches. Alder's cash forecast needs weekly grain (payroll weeks and month-end AP runs create crunches monthly averages hide). Data Foundations drilled this: declare the grain in a sentence before touching data.
• Horizon — the 13-week cash forecast (finance's classic) is a different animal per week: weeks 1-2 are mostly known (booked AR, scheduled AP), weeks 3-6 are pattern-driven, weeks 7-13 are assumption-driven. Good forecast design says which mechanism drives which zone — and Module 3's lab structures exactly that way.