Demand Forecasting

Predicting where demand moves next, county by county.

A nationwide retail-movement forecasting model — county-granular, channel-aware, with a calibrated baseline and a residual correction that delivers sub-2.0 MAE across more than fifty thousand tracked accounts.

Counties tracked2,142
State coverage50/50
Retail channels5
Calibrated MAE1.87

The model knows where it’s confident — and where it isn’t.

A national rollup hides as much as it reveals. County-level forecasting surfaces where growth is real, where it’s softening, and which accounts deserve a rep’s next visit.

Calibration cut error by ~24%

A national monthly multiplier applied on top of the raw model pulled MAE from the pre-calibration baseline down to 1.87. A second-pass residual correction takes it lower still — the model knows its own systematic miss.

County beats state rollups

Forecasting at the county level surfaces growth pockets a state-level model averages away. Across 2,142 active counties, neighboring counties routinely diverge by 10+ percentage points of YoY movement.

Channel-aware MAE varies 2×

Per-channel error ranges from 1.62 in the tightest slice to 3.77 in the loosest. Useful both as an honesty signal and as guidance for where additional data would actually move the number.

38% of accounts flag as at-risk

Roughly 19,000 accounts have a forward forecast that sits below their most recent actual movement. That’s a prioritized call-list for sales without any extra modeling work — it falls out of the same forecast.

From historical movement to a forward-looking county map.

A deliberate pipeline. Each step is a choice about what to model and what to leave to a follow-up data investment.

01

Pull movement history

Twelve months of account-level movement across every modeled county and channel.

02

Model per channel

Each retail channel gets its own baseline — high-frequency channels behave nothing like long-tail ones.

03

Calibrate and correct

National monthly multiplier, then a residual correction pass to pull MAE down further.

04

Surface county deltas

Three months forward, ranked by predicted movement vs current actuals.

Where the next data dollar moves the number.

The model knows its own gaps. The data investments below were scoped against actual MAE deltas observed when each signal was introduced to a slice of the model.

Scenario 01

Syndicated POS data

Third-party point-of-sale across a set of priority markets, monthly refresh.

  • Identifies promotional lift patterns the model currently misses
  • Reps can prioritize under-stocked promo accounts
  • Closes a known systematic miss in promo-heavy retail
Scenario 02

Hyper-local demand signals

Neighborhood-level affluence, traffic, and visitor signals layered onto each account location.

  • Scores accounts in the weakest-MAE channel by local growth potential
  • MAE drops ~20% in modeled slices of that channel
  • Field reps focus visits on high-potential locations
Scenario 03

Real-time channel telemetry

Daily movement feeds from channel partners — closing the gap between actuals and the next forecast pass.

  • Latency drops from 30 days to 1 day
  • Catches demand shifts within days, not months
  • Enables cold-start forecasting for new SKUs

Forecast error varies more than 2× across channels.

Per-channel MAE tells you both how much to trust each slice of the forecast and where the next data investment is most likely to move the number.

Mean absolute error by channel

lower is better · per account-month
Channel A
1.62
Channel B
2.03
Channel C
2.44
Channel D
2.81
Channel E
3.77
Model: per-channel baseline + national calibration + residual correction Forward window: 3 months · county-granular