Baseline comparison
A test of simple rules, standard statistical methods, and more complex models on your data, so complexity is added only when it helps.
Data science and predictive analytics
We build forecasting and classification models for planning and operations, and test them against simpler methods on data held back from development.
ScopeA single pipeline, report, or model, up to a set of connected sources
Delivery timingWeeks for a well-defined first release when data access and definitions are agreed
How we check itEvery model is tested against simpler methods on data it has not seen.
Test whether prediction improves the decision
A prediction is only worth building if it improves a decision. We start by defining that decision, how far ahead it must be made, and what different errors cost. That tells us how accurate a model has to be before it is worth deploying.
Good fit
- Teams planning demand, capacity, revenue, or inventory
- Operations teams that need earlier warning of problems
- Software products with enough history to make repeatable predictions
Not the right fit
- Predictions no one is responsible for acting on
- Organizations whose historical data is too unreliable to learn from
- Problems a simple rule already handles well
What you get
Validated model
A tested forecasting or classification model, with its inputs, uncertainty, and typical errors documented.
Deployment plan
How predictions reach the people or systems that use them, how they are monitored, and when to review, retrain, or retire the model.
How it works
Start with the decision
With the decision owner, we define how far ahead to predict, what action follows, and what each kind of error costs.
Test on held-out history
We test on past data the model never saw, check for leakage (information it would not have at prediction time), and compare results across segments.
Pilot the response
Your team uses the output on a real decision, and we check whether it helped before its role expands.
Scope, cost, and ownership
What we need from you
- Relevant historical records and known outcomes
- An owner who can define acceptable errors and act on results
What affects cost
- Data preparation and labeling
- Model complexity and depth of testing
- How quickly predictions are needed, and monitoring requirements
Technical scope
- Forecasting
- Anomaly detection
- Classification
- Feature definition (the inputs a model uses)
- Backtesting on past periods
- Model monitoring
Support and maintenance
The support plan sets out when to review prediction errors, adjust thresholds, and retrain or retire the model, and who makes those decisions.
Common questions
Do we need machine learning?
Not always. A simple rule or statistical method is sometimes enough, and we recommend the least complicated approach that works.
Can a forecast be explained?
Yes, to a degree that depends on the method. We show the inputs, assumptions, and uncertainty, and compare the forecast with past outcomes.