Predictive Analytics
A forecast is an informed estimate, not a fact about the future. Its value lies in being wrong by less than your current method, and in telling you how wrong it is likely to be.
We build forecasts for three recurring business questions: how much will sell, which customers are about to leave, and which machine is likely to fail. Each is delivered as a range with a most likely value, alongside the record of how earlier forecasts compared with what happened.
Three forecasts businesses ask for
The mathematics overlaps. The data and the pitfalls differ.
Demand
Units per product, per location, per week. Inputs are sales history, price changes, promotions, holidays and sometimes weather. The hard cases are new products with no history and items that sell a few units a month.
Churn
A score for each customer estimating the chance they cancel or stop buying. Inputs are usage, support contacts, payment behaviour and contract dates. A score is only useful if someone then phones the customer.
Maintenance
An estimate of remaining life or failure risk for a machine, from sensor readings and service logs. It needs recorded failures to learn from, which well-maintained equipment rarely supplies.
What we say about uncertainty
These statements go into every report we write.
Every figure has a range
Next month is given as a likely value with a low and a high bound. Planning for the range is the point of the exercise.
Further ahead means wider
A forecast for next week is tighter than one for next quarter. We show how the range grows with the horizon.
Breaks are not foreseen
A model built on history cannot anticipate a new competitor, a regulation or a pandemic. When conditions shift, it must be refitted.
Correlation is not cause
The model may find that customers who call support twice tend to leave. That does not show that the calls drive them away.
Sometimes the simple method wins
If a moving average performs as well as a complex model in testing, we recommend the moving average.
How a forecasting project proceeds
Short stages, each with a result you can inspect.
Review the history
How far back it goes, what is missing, and which past events distorted it, such as stock-outs that hid true demand.
Set the benchmark
Your present method, or a naive rule like "same as last year", is measured first.
Back-test
We pretend to stand at past dates, forecast forward, and compare with the known outcome. This is repeated over many dates.
Deliver where you work
Results land in the spreadsheet, dashboard or planning system your team already opens each morning.
Compare every month
Forecast against actual, in a standing report. Persistent error in one direction is investigated.
What we will ask you for
Gathering these before the first call saves time.
- Transaction or event history, as far back as it exists
- Dates of price changes and promotions
- Periods when stock ran out or a site was closed
- Customer start and end dates for churn work
- Sensor logs and service records for maintenance work
- Your current forecast, however rough
- The decision the forecast will inform
- The cost of forecasting too high versus too low
Methods and software
Chosen by back-test result on your data, not by fashion.
- Classical methods
- Exponential smoothingARIMASurvival analysis
- Learned models
- LightGBMXGBoostQuantile regression
- Libraries
- statsmodelsscikit-learnpandas
- Delivery
- Excel or Google SheetsPower BIMetabaseREST API
What buyers want to know
How close will the forecasts be?
We cannot say before testing, and we do not promise a level of accuracy. The back-test produces measured error on your own history, and you decide from that whether to proceed.
We only have one year of history. Is that enough?
For weekly patterns, perhaps. For yearly seasonality, no, because the model has seen each season once. We will tell you what can and cannot be estimated.
Can external data be added?
Yes, where it is available ahead of time. Public holidays are known in advance. Weather beyond several days is itself a forecast and adds its own error.
Who acts on the predictions?
Your people. A buyer sets the order quantity and an account manager calls the customer. The forecast informs them and does not decide.
How often is the model refitted?
On a schedule agreed with you, commonly monthly, and whenever the monthly comparison shows it drifting.
Tell us what you are trying to predict
Say what you forecast today, how you do it, and what it costs when the number is off. A reply reaches you within one business day.