← Archimedes02 / 04
Proprietary engine — module 02

Predictive modelling

A forecast is only useful if it comes with an error bar and an explanation. Ours do.
Predictive modelling — Archimedes module at Atlas Marketing
The forecasting layer combines a time-series baseline with a marketing-mix component. The baseline captures trend, yearly and weekly seasonality, holidays and known one-off events. The marketing component estimates the incremental contribution of each channel using adstock (carryover) and saturation curves, so we can tell the difference between a good month and a heavily funded one.
Models are refit monthly and validated on held-out periods — we report mean absolute percentage error on the last three months rather than on the data the model was trained on. Typical accuracy on a stable retail or hospitality account runs in the 8–14% MAPE range at monthly granularity; new products and volatile categories are wider, and we say so.
Every forecast is delivered as three scenarios — conservative, base and stretch — each with the assumptions written next to it. A number without its assumption is not a forecast, it is a wish.

Specification

Granularity
Weekly forecast, monthly reporting
Horizon
Up to 12 months; 3 months highest confidence
Typical accuracy
8–14% MAPE on stable accounts
Refit cadence
Monthly

Data inputs

  • 24+ months of demand or revenue history
  • Media spend by channel and week
  • Price and promotion calendar
  • Distribution or store-count changes
  • Category seasonality from the market-analysis module

How it runs

4 stages
  • 01

    Assemble

    Daily or weekly demand series joined to spend, price, promotion, weather and calendar covariates.

  • 02

    Fit

    Bayesian time-series baseline plus adstock and Hill saturation curves per channel, with priors set from category benchmarks.

  • 03

    Validate

    Rolling-origin backtesting on held-out windows; MAPE and bias reported per horizon.

  • 04

    Scenario

    Budget, price and launch-date scenarios simulated and returned with credible intervals.

Methods inside the model

4 methods
  • 01

    Bayesian structural time series

    Separates trend, seasonality and event effects, and returns credible intervals rather than a single line.

  • 02

    Adstock & saturation

    Geometric carryover plus Hill curves estimate diminishing returns per channel and the point where extra spend stops working.

  • 03

    Elasticity estimation

    Price and promotion elasticity by product tier, used to test discount decisions before they are made.

  • 04

    Rolling backtests

    Accuracy measured on periods the model never saw, reported honestly alongside every projection.

What it produces

  • /12-month demand forecast with credible intervals
  • /Three-scenario planning view with stated assumptions
  • /Channel response curves and saturation points
  • /Accuracy report (MAPE, bias) refreshed monthly

Honest limits

  • New brands with under 12 months of history are modelled from category analogues and carry much wider intervals.
  • The model cannot predict shocks — supply failures, regulation, a viral crisis. It can quantify their impact afterwards.
Next moduleBudget allocation →
Talk to us about Archimedes →