Introducing DoubleCast

Efficient, context-aided probabilistic forecasting for the real world.

Vincent Zheng, Étienne Marcotte, Arjun Ashok, Andrew Robert Williams, Lijun Sun, Alexandre Drouin, Valentina Zantedeschi

ServiceNow Research, McGill University, Université de Montréal, Mila - Quebec AI Institute

The Problem

Context-aided forecasting integrates external knowledge—incident reports, domain narratives, policy changes—with numerical time series to improve predictions. Yet multimodal models often fail to outperform unimodal baselines. We hypothesize the bottleneck is data quality: many existing datasets rely on contextual sources whose relevance to the forecast and complementarity with historical data is uncertain, including matched web text, generated descriptions, or templated context.

CAF-7M Dataset

We introduce CAF-7M, a semi-synthetic corpus of 7 million context-augmented time series windows spanning 11 domains. Our pipeline generates plausible scenarios using an LLM conditioned on both history and future values, then verifies test-set contexts by checking whether a strong forecaster improves when conditioned on them. This yields a rigorously validated test set with context that are relevant and complementary to numerical histories. The test set contains 904 verified windows split into HARD and EASY subsets.

Data augmentation pipeline: sampling windows, generating context with an LLM, and verifying context improves forecasts — The data-augmentation pipeline: sample forecasting windows, generate scenario-style context, verify that including context improves predictions.

Model Architecture

To showcase the value of CAF-7M, we train a multimodal model on it: DoubleCast. DoubleCast augments the Chronos time series foundation model with a pre-trained text encoder (Qwen3-14B). Each decoder block gains a DualT5 cross-attention layer that attends to text embeddings alongside the standard encoder–decoder attention over time series embeddings. This lightweight alignment largely preserves Chronos's numerical forecasting capability while enabling context-conditional predictions.

DoubleCast architecture: Chronos encoder and Qwen text encoder feeding into DualT5 decoder blocks with dual cross-attention — Each DualT5 decoder block: masked self-attention, Chronos cross-attention, DualT5 text cross-attention, and FFN—with residual connections and layer norm throughout.

Results

Why filtering matters

On the Accepted set, correct context significantly improves both CRPS and win rate across all LLMs and multimodal forecasters. Swapped contexts do not improve over no context, confirming that gains are driven by complementary information—not any random text. On the Rejected set, providing context actually hurts accuracy. This highlights the importance of filtering: without it, low-quality contexts could lead to the wrong conclusion that a method cannot leverage context, when the contexts are simply not useful.

CRPS comparison across multimodal forecasters on Accepted vs Rejected sets — Mean CRPS (lower is better) across multimodal forecasters on the Accepted and Rejected sets.

Win rate against Chronos across multimodal forecasters on Accepted vs Rejected sets — Win rate against Chronos across multimodal forecasters on the Accepted and Rejected sets.

Benchmark

DoubleCast trained on CAF-7M achieves the highest win rate across all splits and is competitive with Direct Prompt using GPT-5.2 in CRPS, ranking second on ALL and HARD while achieving the best CRPS on EASY. Removing or swapping context markedly degrades both metrics, confirming that DoubleCast actively leverages context.

Model	ALL		HARD		EASY
	CRPS	Win	CRPS	Win	CRPS	Win
Chronos	0.278	—	0.304	—	0.247	—
ChatTime	0.385	0.25	0.407	0.25	0.358	0.25
DP Qwen3-14B	0.296	0.51	0.277	0.63	0.319	0.36
DP GPT-5.2	0.217	0.62	0.212	0.75	0.223	0.47
DoubleCast	0.231	0.71	0.251	0.79	0.204	0.64
no context	+18.3% / −34.7%		+22.0% / −41.9%		+11.1% / −21.5%
swapped context	+26.5% / −37.6%		+25.2% / −40.8%		+25.4% / −37.7%

Benchmark on CAF-7M test set (time series are out-of-distribution from training). CRPS (lower is better) and win rate vs Chronos. DP = Direct Prompt, an LLM-based method that prompts a language model to forecast directly from tokenized time series and context [1]. Ablation rows show relative change in CRPS / win rate when context is removed or shuffled.

[1] Williams et al., "Context is Key: A Benchmark for Forecasting with Essential Textual Information," ICML 2025. arXiv:2410.18959