I just wrapped up the forecasting module in my MBA program and wanted to capture what stuck before it fades. This one hit different because so much of it maps directly to what I do at work every day. The module covered three big buckets: time series methods, regression-based forecasting, and judgment-based approaches. Time series was the most familiar: moving averages, exponential smoothing, seasonal decomposition. I've been doing versions of this in Excel for years without knowing the formal names. Turns out my "trailing 12-month average with a seasonal bump" is basically Holt-Winters. Who knew. Regression was where things got interesting. The idea that you can isolate the effect of individual drivers on your forecast, headcount on payroll, pipeline on revenue, usage on cloud costs, feels like it should be standard practice in FP&A. But most teams I've worked on just do top-down growth rates or bottoms-up builds without much statistical rigor. There's a middle ground here that I want to explore more. The biggest surprise was the section on forecast accuracy measurement. We talked about MAPE, bias, and tracking signals. I realized I've been measuring forecast accuracy inconsistently: sometimes looking at absolute error, sometimes directional, sometimes just vibes. Having a framework for this matters more than I thought. It's not just about being right, it's about understanding *how* you're wrong so you can get better. One connection I keep making: the variance analysis work I've been building ties directly into this. If you can decompose a variance into price, volume, and mix, you're basically doing a post-mortem on your forecast. The [[lab/Variance Analysis Bot|Variance analysis bot]] I've been prototyping could eventually feed back into forecast models. And the variance analysis framework I've been building already captures some of this thinking, though I want to update it with what I learned here. Next step is to try building a simple regression-based forecast for one of my cost centers and see if it beats the spreadsheet method. Small experiment, low stakes. And yes, I know there's a decent chance leadership ignores whatever model I build, which is [[signal/Seven Truths About FP&A|a whole separate reality of this job]]. ← [[Notebook|Back to /notebook]]