Introducing ci-proto.
Overview
ci-proto is out today. It’s the debut model family from Continuous Intelligence and our first step towards a new generation of time series foundation models. It’s available to use now.
fev-bench has a new #1 foundation model: ci-proto XL, the largest in our family at 167M parameters and still only a prototype. The smaller models deliver too. The five ci-proto models form the Pareto frontier for forecast error against model size and runtime. No other public time series foundation model forecasts more accurately for the same size or speed.
ci-proto also sees further back. It can recall patterns from more than 2 million observations ago, with a context window 64 times larger than any other time series foundation model’s.
This launch delivers the first element of our novel approach to time series. Next, we’re moving on to bigger and better models with entirely new capabilities for real-world data.
What is a time series foundation model?
Time series data sit at the heart of some of the world’s most important systems, from machine sensors and trading feeds to electricity grids and patient vitals. The forecasts and insights drawn from such data inform critical decisions, such as when to maintain equipment, how to manage financial risk, and when to charge or discharge grid batteries. But creating solutions that can make sense of such complex data takes specialist expertise, and each new application or question can require those solutions to be adapted.
Previous approaches leave significant gaps. Statistical models like ARIMA and VAR need to be fitted separately to each dataset and rely on linear relationships, limiting their ability to capture more complex patterns. Machine learning models can learn richer patterns across many variables, but each new dataset or application often needs a bespoke model, trained on large amounts of task-specific data.
Time series foundation models (TSFMs) take a different route. They learn from many datasets during pretraining, the approach behind the success of large language models. One model can then tackle many applications, drawing on patterns learned in different domains without starting from scratch each time. Our TSFM, ci-proto, can forecast on unseen datasets straight away without extra training, an ability known as zero-shot forecasting.
Try it yourselfLeading benchmark results
ci-proto is the #1 foundation model on fev-bench, a key benchmark for multivariate forecasting with time series foundation models. Furthermore, ci-proto is also available for commercial use.
fev-bench leaderboard
| Rank | Model | Skill score (%) |
|---|---|---|
| #1 | ci-proto XL | 51.5 |
| #2 | TimeH2O-v1 | 50.8 |
| #3 | Timestral-209M | 50.4 |
| #4 | t0-gamma | 49.4 |
| #5 | TimesFM-3 | 49.0 |
| #6 | Chronos-2 | 47.3 |
On fev-bench, the ci-proto family sets a new frontier for forecasting error against both parameter count and runtime. In other words, ci-proto is the better choice, whatever your choice of forecasting speed or model size.
fev-bench pareto plots
Scaled quantile loss (SQL), averaged across tasks.
Scroll horizontally within the plot to read every label.
Scaled quantile loss (arithmetic mean across tasks). Lower is better on both axes.
The horizontal axis is logarithmic.
Goodbye context windows
Most existing time series foundation models use Transformers to process history. Attention enables accurate recall, but its computational cost grows quadratically as the number of past observations increases. With ci-proto’s novel attention-free time mixer, the computation needed to read history only grows linearly with its length. This makes much longer context windows practical. We show that ci-proto can use information from a context window 64 times as long as the context window of Timestral (Mistral).
How computation scales with context length: ci-proto vs Transformers
Illustrative growth rates. Each square represents one unit of each method’s baseline computation. Model size, patch size, and the number of variables stay fixed.
With a long context window, ci-proto can draw on historical data better than any other TSFM, for example, in a scheduled power outage.
With 4 years of history, the earlier scheduled power outage is in context. The forecast captures the drop in electricity use.
- Electricity use
- Median forecast
- 10–90% forecast range
One eight-hour window from this experiment, using the ci-proto long-context checkpoint. Both forecasts receive the same future power outage schedule.
Get started with forecasting using ci-proto
Run ci-proto locally on your data, on CPU or GPU, with our Python package. Supply past observations and any known future inputs, such as operating schedules. The model returns 9 forecast quantiles for each target variable to describe a range of possible outcomes.
pip install continuousintelligenceWhat’s next for Continuous Intelligence?
ci-proto is our first public model, and frontier forecasting performance on existing benchmarks is just our starting point. Join our newsletter and waitlist to stay informed about the latest developments in time series and our upcoming releases.