Cellular capacity has to be bought before it is needed, because equipment takes months to arrive and install. This forecasts traffic six or more months ahead, from individual sectors up to the whole network, so the spending goes where the demand is going to be.
You buy capacity before you can see the demand
Capacity planning fails in two directions and both are expensive. Under-build and the network congests where the growth actually happened, which customers experience as the product being bad. Over-build and money sits in equipment serving demand that never arrived. Extrapolating last year forward does not help, because cellular traffic is strongly seasonal and the seasonality is not the same shape in a business district as it is at a stadium or on a motorway.
Forecasting at more than one altitude
A regional total and an individual sector are different problems even though they are the same measurement. Aggregate traffic is smooth and forecasts well. A single cell is spiky, and its history includes changes that were not demand at all - a neighbouring site coming online, a re-parameterisation, a sector being re-pointed. So forecasts run from individual sectors and cells up through regions to the network, and the models are aware of the network configuration behind the series rather than reading every jump as a change in user behaviour.
More than one model, on purpose
SARIMA and Prophet handle seasonality differently and fail differently: one is more disciplined about the statistical structure, the other more forgiving of missing data and abrupt level shifts. Running both, plus other approaches, is not indecision. Where two methods with different assumptions agree, the forecast is doing something more than fitting the quirks of one method, and where they diverge, that divergence is the interesting signal about the series rather than a nuisance.
What comes out
Six-month-plus forecasts with the seasonal structure separated out, at every level from sector to network, and recommendations about where capacity and resources should go. The horizon is not arbitrary: it is set by how long it takes to procure and install, because a forecast shorter than the lead time cannot change a decision.
The forecast that is easy to produce and impossible to spend
Both shortcuts produce a number. Neither produces a purchase order anyone will sign.
- Forecast only the sectors already under strain
- Modelling the sectors that congest today is the cheap version and it forecasts the wrong ones. A sector that is struggling now is one the planners already know about; the value is in the sector that looks comfortable and will not be in eight months, which is exactly the one a shortlist built from today's alarms leaves out.
- One model for every sector
- A single model applied everywhere is defensible right up to the point where you look at the seasonality. A sector at a university, a beach and a business district do not share a shape of year, and a model fitted to all three learns the average of three seasons that no sector has.
The deliverable is an argument, not a model
The deliverable is not a model. It is an argument about where money should go next, with the reasoning exposed so somebody can disagree with a specific part of it. The opening days of a build produce that same argument, about processes rather than about base stations.
What it proves
Capacity planning backed by data: 6+ month forecasts that show where the capacity investment is needed before the money goes out.
Category
Consulting
Built with
Time Series Forecasting · SARIMA · Prophet · Machine Learning · Network Analytics · Data Science
Need something similar?
The cheapest way in is two weeks. The first days work out which task would pay for itself in your processes; the rest builds that agent on your own data and measures it.