DoubleCast

Code for the DoubleCast model architecture for context-aided time series forecasting, introduced in Overcoming the Modality Gap in Context-Aided Forecasting.

This model was trained using the CAF-7M semi-synthetic dataset, also introduced in the same paper. The trained weighted are available at HuggingFace.

A comparison of forecasts between DoubleCast and Chronos can be seen in the associated demo.

Disclaimer: this code is not compatible with newer versions of the transformers library. This version is only known to be compatible with version 4.51.3 of said library.

Usage example

import torch
from functools import partial
import pandas as pd
import numpy as np

from doublecast.models.doublecast import DoubleCastPipeline
from doublecast.data.prompts import prompt_structured

# Load from HuggingFace Hub (downloads delta weights only;
# base Chronos and Qwen3-14B weights are fetched from their original repos)
pipeline = DoubleCastPipeline.from_pretrained(
    "ServiceNow/DoubleCast",
    device_map="cpu",
    torch_dtype=torch.bfloat16,
)

past_timestamps = pd.date_range(start="2023-01-01", periods=20, freq="D")
future_timestamps = pd.date_range(start="2023-01-21", periods=10, freq="D")
freq = "D"
past_values = np.random.rand(20) + 0.2 * np.arange(20)
text_context = "The past values show a clear upward trend, with some fluctuations. However, afterwards a massive drop is expected in the values."

prompt_func = partial(
    prompt_structured,
    past_timestamps=past_timestamps,
    future_timestamps=future_timestamps,
    freq=freq
)

context = torch.tensor(
    past_values,
    dtype=torch.bfloat16,
).flatten()

dual_preds = pipeline.predict(
    context=context,
    text_context=text_context,
    prediction_length=len(future_timestamps),
    num_samples=20,
    limit_prediction_length=False,
    format_prompt=prompt_func,
)