Time series
Demand, load or prices, alone or for thousands of related series together, with promotions, weather and holidays as inputs.
Tables
One row per customer, loan or transaction.
Event histories
Entities with any number of timestamped records. An example is a borrower with several accounts, each with 36 months of payment history.
Relational data
Several related tables, such as customers, orders and products, read as they stood at the time of each prediction.
The problem
Most prediction projects spend their time before any model is fitted.
- Data has to be joined and cleaned.
- Features have to be designed by hand: lags, rolling averages, summaries of past events.
- A model is then fitted and tuned for one question.
- A new question or a change in the data means doing much of it again.
Many firms have no team for this, so decisions run on rules of thumb and spreadsheets.
What tabSutra does
- You pass your tables with their timestamps.
- The model reads the past examples in that data and predicts for the new cases.
- Results come in hours, not months.
- With enough history, the model is fine-tuned on your data for higher accuracy.
Data that comes as sequences, such as the months of a loan account, is read as a sequence. A standard model can only see such data as separate rows or as hand-made summaries.
What you can predict
Forecasts
Forecasts with ranges, so you plan for the likely case and the bad case.
Probabilities
Probabilities, such as default, churn or fraud.
Amounts and durations
Amounts and durations, such as loss given default or time to failure.
Two levels at once
Predictions at two levels at once, such as each account and the customer, or each SKU and the store.
New products, new stores, new markets
- A new product has no sales history. Neither does a new store or a new category.
- Most forecasting methods need that history, so these cases are planned by hand.
- tabSutra represents every product, store and customer as an embedding.
Forecasts from day one
A new product is forecast from the history of products like it.
Slow movers are not left out
An item with little history borrows from similar items with a lot of history.
Any number of items
The approach works the same for a hundred SKUs or ten million.
Expansion
A business entering a new region or category starts from what it already knows about similar regions and categories.
Use cases
Retail and consumer goods
Demand at SKU and store level, including new launches and new stores. Replenishment and promotion effects.
Lending and financial services
Credit scoring from account histories, early warning, collections prioritisation.
Energy and utilities
Load and generation forecasts, at any time granularity.
Electric vehicle fleets
Battery management systems report voltage, temperature, charge cycles and faults for every vehicle. tabSutra reads these histories across the whole fleet to predict downtime, battery health and residual value.
Manufacturing
Sensor monitoring, failure prediction, maintenance planning.
Built for production data
No look-ahead
A prediction uses only what was known at its time, so back-tests reflect real use.
Missing values are handled as missing
Nothing is filled in by guesswork.
Uneven timestamps are fine
Records can arrive at any time, and late arrivals are treated as late.
Consistent outputs
The same data gives the same prediction. Reordering columns changes nothing.
Ranges, not only point estimates
Every forecast comes with its uncertainty.
How it is used
-
tabSutra is first run on your historical data and compared with your current method.
-
It is then used through an API, or deployed on your own infrastructure so that data stays in-house.
-
Either way, it can be fine-tuned on your full history.
Team
We are a team of applied ML scientists. We have worked on real-life data across industries, with methods ranging from classical machine learning to transformers.
The name
Sutra is Sanskrit for a thread, and for a short rule that holds a body of knowledge. tabSutra is one set of rules for structured data.