Sales, demand and operational forecasting with machine learning.
AI Forecasting and Predictive Analytics Services
Supertype builds forecasting and predictive models that tell your team what is likely to happen next, and how sure the model is, from next quarter's sales to next week's river levels.
Our predictive models run in production for enterprises in mining and logistics, where they save millions of dollars a year in barging costs.
Machine Learning by Skilled Practitioners
Supertype's engineers are exclusively from a Mathematics and Computer Science background, skilled in statistical modeling and machine learning for high-stakes decision making.
Before any model goes live we measure how accurate your current method is, backtest the new one on periods it has not seen, and show the difference. Since 2021, we have worked with 500 corporate clients across various industries in building and deploying predictive models.
Supertype helps companies make better forecasts grounded on advanced modelling techniques and rigorous statistical analysis.
What we forecast
Sales and revenue forecasting
Forecasts by product, customer segment, region or channel, built from your sales history plus the things that move it: pricing, promotions, seasonality and holidays. The hardest part is usually separating a real trend from a one-off promotion, so we model them separately.
What you get
- Forecasts at the level you set targets
- Effects of pricing, promotions and holidays
- A range around every number
Demand and inventory forecasting
Demand forecasts at the level you plan stock, so purchasing and production can order the right amount. The aim is fewer stockouts and less excess inventory, and the range matters as much as the number, because it sets how much safety stock to hold.
What you get
- Demand by item, location and period
- Ranges that feed safety-stock decisions
- Delivery into your planning tools or ERP
Operational forecasting
Models that predict the conditions your operations depend on, such as river levels, volumes, workloads and lead times. They run on live data and are delivered as dashboards and alerts, so the people making the call see the forecast as it changes.
What you get
- Real-time inference on sensor and system data
- Dashboards and alerts for the operations team
- Monitoring for sensors that drift or drop out
Scoring and early warning
Models that score records by the likelihood of an outcome, such as a late payment, a lost customer or an unusual transaction, so your team can act first. A score is only useful if people know what to do with it, so we agree the action and the threshold up front.
What you get
- A score per record, refreshed on a schedule
- Thresholds tied to a defined action
- Anomaly alerts for unusual transactions
How we build a forecasting model
- 01
Decision and data review
We start from the decision the forecast feeds, such as how much to order or produce, then review the history, granularity and quality of the data behind it.
- 02
Baseline
We measure how accurate your current method is, whether that is a spreadsheet, an ERP forecast or a planner's judgment, so every model has something real to beat.
- 03
Features and external drivers
We add the factors that explain demand beyond its own history, such as prices, promotions, holidays, weather and market data.
- 04
Modeling and backtesting
We compare statistical and machine learning models on past periods they have not seen, and pick the one that is most accurate where it matters to your business.
- 05
Delivery
Forecasts land where your team works: dashboards, alerts, reports or your ERP, with a range around each number instead of a single guess.
- 06
Monitoring and retraining
We track forecast accuracy after launch and retrain the models as your business and data change.
Forecasting software or a custom model?
Not every company needs a custom model. Software suits standard planning on clean data. A custom model pays off when the forecast depends on drivers the software cannot use, or the decision is worth a lot. Our build, buy or partner guide covers how to weigh the cost of each over several years.
| Spreadsheet or ERP forecast | Forecasting software | Custom forecasting model | |
|---|---|---|---|
| Best for | Few products, stable demand | Common planning needs with standard data | Many drivers, unusual data or decisions worth a lot of money |
| Uses external drivers | Rarely | Some, within the product | Any data you have or can buy |
| Accuracy you can verify | Usually not measured | Reported by the vendor | Backtested on your own history |
| Fits your workflow | Manual | Within the software | Built into your dashboards, ERP and alerts |
Forecasts inside your ERP
Forecasts are most useful where decisions are made. For companies on Odoo, Superlative answers questions about sales, stock and revenue from your ERP data, and produces forecasts on request, labelled as estimates with every step of the calculation shown.
Frequently Asked Questions
What is predictive analytics?
Predictive analytics uses historical data and machine learning to estimate what is likely to happen next: how much you will sell, which invoice will be paid late, or which customer may stop buying.
The output is a forecast or a score that helps a team decide earlier and with more confidence. We have built these for clients such as Alamtri, where a river-level forecast covers 7 days ahead, and we run them as part of our wider data science consulting work.
What is the difference between sales forecasting and demand forecasting?
Sales forecasting estimates what you will sell and earn, and is mostly used for revenue planning and targets. Demand forecasting estimates how much customers will want, and is mostly used for purchasing, production and inventory.
They often share data and models, but are planned at different levels of detail. A sales forecast might be one number per region per month, while a demand forecast may need one per product and location. We agree which of the two you need before we touch the data.
How much historical data do we need for AI forecasting?
Usually at least two years, so the model can see seasonal patterns more than once. Less can still work for short-term forecasts or when strong external drivers are available.
We check what your data supports before recommending a model. If the history is thin or has gaps, we say so up front and tell you what to start collecting, rather than fitting a model that looks good on paper.
How accurate are AI forecasts?
It depends on how predictable the thing you forecast is, so we do not promise a number. We measure your current accuracy first, backtest each model on past periods it has not seen, and show the improvement before anything goes live.
Every forecast comes with a range, not just a single number, so planners can see how much to trust it. For a financial example, see our financial data science work.
Can forecasts include external factors like holidays, promotions or weather?
Yes. External drivers are often where a model beats a spreadsheet, because a spreadsheet trend line cannot see a public holiday or a price change coming.
We add the ones that matter for your business, such as holiday calendars, promotions, prices, weather and market data.
Can forecasts connect to our ERP or BI tools?
Yes. We deliver forecasts to dashboards, BI tools, scheduled reports and ERP systems, so planners use them in the tools they already work in.
If you run Odoo, we have a dedicated service for AI in Odoo. Where the output is a recurring document, we can also use report automation to send it out on a schedule.
Should we buy forecasting software or build a custom model?
Software is a good fit when your planning needs are standard and your data is clean. A custom model makes sense when the forecast depends on drivers the software cannot use, your data is unusual, or the decisions it supports are worth a lot of money.
The same question comes up for most AI projects, and our build vs buy vs partner guide walks through how to decide.
How much does a forecasting project cost?
Cost depends on how many things you forecast and at what level of detail, how many data sources must be combined, and where the forecasts must be delivered.
We scope these first and send a fixed proposal, so you know the number before work starts. Contact us with a short description of what you want to forecast and we will tell you what we would need to see.
Tell us what you need to predict
Describe how you forecast today and what data you have. We will tell you whether a model would do better, and by how much we would expect it to.
What to bring
The thing you forecast, how you do it today and a sample of the history.
What you get back
An honest view of whether a model beats your current method, and a fixed proposal if it does.