AI Forecasting for Real Business Decisions
Nine live models to play with — from pulling demand apart to the production trust gate.
Hosted by Tim Löhr · Data Engineer at Meta
Time-series structure
Every demand line is really three things added together — separate them before you model.
Read any series as trend + seasonality + noise, and know which parts a model can actually predict.
You model the trend and season; the noise you can only buffer for — chasing it is overfitting.
The foundation for every slide after — next we add signal, choose a level, then measure what's left.
Exogenous regressors
The biggest accuracy win is often telling the model what you already know is coming.
Add known-future events — promo, price, launch, holiday — as inputs and watch error fall.
A driver only helps if it's known at forecast time — otherwise it's leakage, or noise you can't use, like weather.
Decomposition read the past; drivers inject the known future onto the same series.
Where you forecast is a choice
The same demand exists at many levels — SKU, family, region, total.
See why higher levels forecast more accurately — and why you must still act at the SKU.
Bottom-up keeps detail but is noisy; top-down is stable but its fixed share is wrong lower down.
Drivers worked on one series; here we choose which level of the tree to trust.
Industry workflow
Daily, weekly or monthly isn't given — it's a choice.
See how the grain changes the data shape, the model that fits, and the action it drives.
Coarser is smoother and lower % error, but you lose the within-period timing you act on.
Hierarchy aggregated across products; cadence aggregates across time.
Evaluation
A forecast's error only counts if it's measured on data the model never saw.
Hold out unseen time and see how one lucky split can flatter a weak model.
A single split can lie; rolling over many windows gives an honest, stable read.
This is how you'd fairly test every choice from the earlier slides.
Evaluation
One set of errors can produce several different verdicts.
See how MAE, RMSE and MAPE disagree, and pick the one that matches your cost of error.
RMSE punishes big misses; MAPE explodes on small demand — there's no universal best.
Backtest is when you measure; metrics is what you measure.
Uncertainty and buffers
A forecast is a range, and the range is a business input — not a hedge.
Pick a coverage, reveal reality, and size the buffer to the more expensive error.
Wider bands are safer but cost inventory; the build point moves toward the costlier miss.
This is where a forecast finally becomes a decision — the payoff of everything before.
Explainability
Every model is also a choice about whether you can say why.
Compare a readable glass-box against an opaque black-box on the same series.
The black box edges the error, but you can't audit, override or defend its number.
Metrics scored accuracy; this scores trust — which sometimes matters more.
Model choice
No single method wins everywhere — the champion changes with the data.
Pick two models and a dataset, then switch the data and watch the winner flip.
Don't choose a model in the abstract; choose on your decision and your data.
Closes the loop — the honest workflow of backtest + metrics beats betting on one model.