Retiring the Bottom-Up Forecast You Rebuild Every Month

Every month, FP&A teams rebuild a forecast that's largely the same as last month's, plus judgment. Chase the business units for inputs, aggregate the spreadsheets, reconcile the versions, present it, and start over. It's enormous effort for a forecast that's often no more accurate than a decent statistical model would've produced in seconds.

I'll say the quiet part out loud: a lot of bottom-up forecasting is elaborate anchoring on the prior period dressed up as rigor. That's not a knock on FP&A teams — it's what the process incentivizes.

What ML brings

Machine-learning models forecast revenue and expenses from history plus drivers — seasonality, pipeline, headcount, macro signals, whatever actually predicts your numbers. They pick up patterns a manual process misses and they don't get tired or optimistic in Q4. On stable, high-volume lines, ML forecasts frequently beat the manual bottom-up on accuracy, and they produce it in a fraction of the time.

The metric to hold it to is forecast accuracy — MAPE, or whatever error measure you already trust. And here's the thing people skip: you should backtest the model against your historical manual forecasts. If the model's MAPE isn't better than what your team was already achieving, you don't have a case yet. Sometimes it is dramatically better on some lines and no better on others, and that's fine — use it where it wins.

The honest limits

ML is good at continuity and bad at breaks. It learns from history, so a genuine structural change — a new product line, a pricing overhaul, an acquisition, a recession that doesn't resemble the last one — is exactly where it struggles, because there's no history to learn from. That's precisely where human judgment adds the most value. The right model is human-plus-machine: the model handles the base, humans overlay the things the model can't know.

Watch for bias, in both directions. Models can inherit bias from biased historical data — if sales always sandbagged, the model learns to sandbag. And humans reintroduce bias when they override the model to hit a number they want. Track override frequency and whether overrides improved or worsened accuracy; you'll learn a lot about your own forecasting culture. Often the overrides are the problem.

Explainability matters for adoption. A CFO won't stake a board number on a forecast nobody can explain, so favor models that show which drivers moved the number. "Revenue's up because pipeline in the enterprise segment grew and it historically converts at X" is a sentence a business partner can defend. A black-box point estimate is not.

One practical caution: don't boil the ocean. Start with the lines that are high-volume and pattern-driven — they're where ML wins cleanly — and leave the lumpy, judgment-heavy lines to people for now. The goal isn't to remove judgment from FP&A. It's to stop spending judgment on the parts a model does better, and save it for the calls that genuinely need a human who understands the business.