ci-proto forecasting guide: long context, quantiles and honest backtests
Use Continuous Intelligence's ci-proto for time-series forecasting. Understand tensor shapes, known-future covariates, quantiles, deployment and the backtests that expose leakage.
BY EMILY CARTER · FDEINTERVIEWS EDITORIAL · UPDATED NOVEMBER 15, 2026 · 5 MIN READ
PRACTICE THIS:Practise time-series evaluation ·Study ML fundamentals ·Learn model monitoring
ci-proto forecasts numerical time series from observed history and returns a distribution represented by quantiles. Use it as a pretrained forecasting candidate for demand, energy, capacity or sensor workloads, then compare it with a simple baseline on future periods it could not see. Its long-history capability is interesting when distant events carry useful information. It does not make a larger input automatically better, and it does not remove the need for time-aware validation.
Continuous Intelligence publishes the Python package, a model card and a research paper. This guide separates their release claims from the experiments an FDE should run before connecting a forecast to a business decision.
What makes the architecture different?
The paper combines Mamba-3 processing along time with attention across variables. Temporal recurrence supports work that scales linearly with history length; cross-variable attention lets related signals inform one another. Linear temporal scaling still costs more compute as history grows, and adding variables creates another scaling dimension.
The authors evaluate five sizes on 100 fev-bench tasks and report a quality-efficiency frontier. Those are results under their evaluation protocol, including preprocessing and test-time adaptation. They are a reason to try the model, not evidence that it will beat a seasonal baseline on your warehouse's demand.
The released model card lists 167 million parameters. The package selects the standard checkpoint for histories up to 16,384 observations and a long checkpoint beyond that. Its advertised 2M-plus history refers to time-series steps, not an LLM token allowance. Measure the memory and runtime of the actual number of series and steps you plan to send.
Start with the input contract
A row represents a target variable and a column a time step. Align frequency, timezone and observation times before constructing the tensor. Zero is a measurement; a missing observation is different. The package supports NaN for missing values, but that does not repair an incorrect time axis.
| Input | Shape | Example |
|---|---|---|
| Target history | [V, T] | Hourly demand for V sites |
| Past covariates | [P, T] | Weather observations available before the cutoff |
| Future-known covariates | [K, T + H] | Calendar features or a published schedule |
| Forecast | [V, 9, H] | Nine quantiles for each site and future hour |
These are the current package input and output contracts. Never put a future observed target in a covariate row. Related-series selection also needs to use information available at the forecast origin; picking neighbors by their future similarity is leakage.
Our point-in-time correctness question explains why a dataset can look perfectly tidy and still contain tomorrow's answer.
Run one forecast, then inspect its shape
Use Python 3.10 through 3.14 in a separate environment. Install the reviewed package with python -m pip install continuousintelligence==0.1.1. The first load downloads model files, so allow disk space and network access during preparation. This example pins the model revision reviewed for this article:
import torch
import continuousintelligence as ci
model = ci.load_model(
"ContinuousIntelligence/ci-proto",
revision="ee3031351a151e3f6febf8c62b53e6ea1008bef6",
device="cpu",
)
history = ci.load_example()
forecast = model.forecast(history, horizon=24)
assert forecast.shape == (history.shape[0], len(model.quantiles), 24)
assert torch.isfinite(forecast).all()
median_index = int(torch.argmin(torch.abs(model.quantiles - 0.5)))
median = forecast[:, median_index, :]
print(history.shape, forecast.shape, median.shape)
We ran this CPU example with continuousintelligence==0.1.1 and the pinned revision. The history had shape [1, 120], the forecast [1, 9, 24], and every forecast value was finite. This was an integration check, not a forecast-accuracy or long-context benchmark.
The bundled example uses the UCI Bike Sharing dataset. A shape check verifies the integration contract; it cannot establish forecast quality. To evaluate quality, reserve actual later observations and compare predictions made before those observations became available.
Backtest the decision you intend to make
For hourly staffing, choose several past forecast origins. At each origin, reconstruct only the history and schedules known then, predict the next 24 hours and score against later observations. Move the cutoff forward and repeat. Keep a final time period untouched while choosing context length and configuration.
In the figure, each blue block stops before its green evaluation interval. The amber calendar row can cross the cutoff only because those values were already known.
Start with a seasonal-naive baseline, such as the same hour last week when weekly seasonality is appropriate. Report median absolute error alongside quantile loss and observed interval coverage. Compare like horizons: a model good one hour ahead can be poor a day ahead. Group results by site, peak hours and missing-data severity so easy periods cannot hide operational failures.
For quantile q, pinball loss is max(q * error, (q - 1) * error), where error = actual - prediction. Underprediction at a high quantile is penalized more heavily. A nominal 10th-to-90th band should be checked for coverage and width together: an enormous band can cover the outcome while being useless for planning. See calibration and uncertainty and time-series fundamentals.
Turn quantiles into an operating choice
| Use case | Forecast to evaluate | Decision still owned by the application |
|---|---|---|
| Inventory | Demand over replenishment lead time | Order size under shortage and holding costs |
| Compute capacity | Load over provisioning delay | Replicas within quota and spend limits |
| Building operations | Energy demand by interval | A bounded control plan with operator override |
| Maintenance | Expected sensor range | Investigation when observations depart from it |
An unusually large residual is a signal to investigate, not a diagnosis of equipment failure. Forecasts of individual sites also need reconciliation if a parent total must equal their sum. Independently taking each site's upper quantile does not produce the upper quantile of aggregate demand.
For a proposed capacity controller, run in shadow mode first: record what it would allocate without changing production. Compare service violations and unused capacity with the current policy. This connects forecasting to model monitoring, where delayed outcomes matter more than a persuasive chart.
Host and adapt it deliberately
Prepare a pinned model bundle once, then load it into a long-lived worker. Bound history size, variable count and horizon at the service interface. Separate cold-start time from warm prediction time, and retain a fallback forecast when inputs are late or inference misses its deadline. Test CPU first for a small workload; evaluate GPU execution on the intended batch size.
The release exposes preprocessing and calibration settings in tta.json. Preserve that file with the checkpoint revision. Test-time adaptation is distinct from updating model weights; do not describe the quickstart as fine-tuning. Begin improvements with input alignment, lookback ablations and covariate availability before attempting a training procedure the quickstart does not supply.
Finally, review the ci-proto Code and Model Licence. It includes attribution and notification requirements beyond Apache 2.0 and distinguishes internal from external use. Record the chosen model, configuration, validation window and release criteria together, as in our model registry concept.
Turn it into offers. Work the real questions and concepts this maps to:
FAQ
ci-proto is Continuous Intelligence's pretrained model for probabilistic time-series forecasting. The released package forecasts one or more target series and returns nine quantiles per target and future step, with optional covariates and related-series context.
The paper describes a time-series architecture that combines Mamba-3 temporal processing with attention across variables. It forecasts numerical sequences directly. A language model may explain a forecast, but it is not the forecasting mechanism described here.
No. Its code and model use a custom license based on Apache 2.0 with additional attribution and notification terms. Read Section 10 before distributing it or offering an external product that uses it; the license distinguishes purely internal use.
The documented input history is a two-dimensional tensor shaped [variables, time]. Forecasts have shape [variables, quantiles, horizon] and are returned as CPU float32 tensors in the target's original units.
No. It is a nominal central 80 percent prediction interval. Measure observed coverage on later data, by forecast horizon and important segments, before using it for capacity or inventory decisions.
Discussion (3)
Future covariates need an availability timestamp. Tomorrow's published staffing schedule is available now; tomorrow's realized staffing is not. Using the second in a historical backtest measures a system that could not have existed.
A long history should earn its place through an ablation. Compare the same forecast origins with several lookback windows, and include the time spent preparing and transferring each input.
Forecasting returns an uncertain estimate. A capacity controller still needs a fallback when the forecast is late, the inputs are incomplete or the proposed allocation exceeds a hard operating limit.
