Model Card for DeOSAlphaTimeGPTPredictor-2025

Model Details

DeOSAlphaTimeGPTPredictor is a multivariate, probabilistic time-series foundation model that supports optional covariates and integrates an ensemble of diverse foundation model paradigms. Trained on close to 1T financial data points across multiple frequencies, it achieves strong generalization performance although it has only seen financial data.

During inference, it leverages in-context learning to further enhance predictive accuracy. The model incorporates a neurosymbolic controller that reasons over the conceptual and structural constraints inherent in the target series. Its primary objective is to deliver accurate, uncertainty-aware forecasts designed to augment and inform decision-making processes.

State of Research

A detailed paper outlining the model architecture, implementation, and experimental results is currently under peer review. Upon publication, the accompanying GIFT-EVAL evaluation script will be released to the public to facilitate reproducibility and further research.

Example Usage

# Requires: deosalpha >= 2025.10.12
from deos_alpha.infrastructure.environment_reader import EnvironmentReader
from deos_alpha.capabilities.forecasting.predictors.forecasting_predictor import (
    ForecastingPredictorConfig,
    DeOSAlphaTimeGPTPredictor,
)

# Initialize configuration for the forecasting model
config = ForecastingPredictorConfig(
    model_id="DeOSAlphaTimeGPTPredictor-2025",
    prediction_length=252,  # Number of time steps to forecast
    api_key = EnvironmentReader().get_deosalpha_api_key()
)

# Instantiate the predictor
predictor = DeOSAlphaTimeGPTPredictor(config=config)

# Fit and predict using in-context learning capabilities
# - `measure_types`: a list of optional strings describing the target concepts to forecast.
#   These inform the neurosymbolic engine, which enforces conceptual and physical
#   constraints (e.g., valid ranges, monotonicity, or bounded behavior).
# - `data`: a DataFrame or dict containing the time series inputs, including
#   `context_id`, `date`, `target`, and optional covariates.
# - `num_samples`: number of probabilistic samples for uncertainty estimation.
# - `frequency`: inferred automatically when set to "auto".
predictions = predictor.fit_predict(
    data=data,
    measure_types=measure_types,
    num_samples=500,
    frequency="auto",
)
#alternatively direct prediction
predictions = predictor.predict(
    data=data,
    measure_types=measure_types,
    num_samples=500,
    frequency="auto",
)