Chrona: A Multimodal Probabilistic Time Series Foundation Model Enabling Event Understanding and Uncertainty Quantification in Predictions

Chrona is a multimodal time series foundation model integrating Transformer and Mamba architectures. It aims to address the limitations of traditional prediction models: being blind to events, struggling to integrate multimodal information, and failing to quantify prediction uncertainty. Its core capabilities include multivariate and hierarchical prediction, event condition injection, probabilistic output, ultra-long context support, in-context learning, and edge deployment, providing new solutions for fields like financial risk control and energy dispatching.

Time series prediction is crucial in key fields like financial risk control and energy dispatching, but traditional models have obvious shortcomings:

1. Being 'blind' to real-world events (e.g., interest rate hike announcements, Black Friday) and unable to integrate multimodal information like text;
2. Point estimate outputs cannot quantify uncertainty, which is a fatal flaw in risk-sensitive scenarios;
3. Poor performance when black swan events occur suddenly.

Chrona uses an 8-layer hybrid backbone network: even layers use Transformer blocks to achieve global cross-sequence attention, while odd layers use Mamba blocks to efficiently capture long-range dependencies, balancing complex correlation modeling and ultra-long sequence processing efficiency. The input-side multimodal encoder integrates three types of information: multivariate time series, external covariates (e.g., holidays), events and text; the output side uses a mixture density network to generate a complete probability distribution and provide quantile predictions.

Event-Aware Prediction

Can input event lists (e.g., 'Black Friday'), learn the impact patterns of events on historical data and apply them to future predictions, suitable for scenarios like retail promotions and financial policy analysis.

Probabilistic Prediction and Uncertainty Quantification

Provides P10/P50/P90 quantile predictions, helping decision-makers understand pessimistic, most likely, and optimistic scenarios, and guiding risk-sensitive scenarios like inventory management.

In-Context Learning and Zero-Shot Adaptation

No retraining needed; inputting a few examples can adapt to new domains, reducing deployment costs.

Edge Deployment

Supports ONNX export (2-5x CPU inference acceleration), Docker containers, and FastAPI services, adapting to edge devices.

API Functions Cover the Entire Workflow

Endpoint	Function	Typical Scenario
/forecast	Probabilistic multi-step prediction	Sales prediction, demand planning
/simulate	Scenario analysis	What-if analysis
/anomaly	Anomaly detection	Equipment failure warning
/embed	Text embedding	Convert event descriptions to model inputs
/forecast/stream	Streaming prediction	Real-time monitoring

Comparison with Traditional Solutions

Capability	Chrona	Traditional statistical models	Classic deep learning models
Multivariate + hierarchical	✅ Natively supported	⚠️ Requires special handling	⚠️ Partially supported
Text/event conditions	✅ Built-in	❌ Not supported	❌ Not supported
Probabilistic output	✅ Natively	⚠️ Approximate	⚠️ Requires additional modeling
4K+ context	✅ Supported	❌ Not supported	❌ Limited by architecture
In-context learning	✅ Supported	❌ Not supported	❌ Requires fine-tuning
Edge deployment	✅ ONNX export	✅ Lightweight	⚠️ Complex

Chrona represents a new direction in time series prediction: shifting from function fitting to intelligent prediction that 'understands events, quantifies uncertainty, and adapts to scenarios'. For data science teams, its event-aware capability understands business context, probabilistic output supports risk assessment, and in-context learning enables rapid deployment. In the future, we look forward to more specialized foundation models like Chrona, bringing out-of-the-box intelligent capabilities to vertical domains.