Forecasting

GradientCastFM

State-of-the-art 0-shot time series forecasting powered by transformer-based foundational models. Predict future values across any domain without training.

Get API KeyView Examples

What is GradientCastFM?

GradientCastFM is a foundational model for time series forecasting. Built on the same transformer architecture that powers large language models, it has been trained on billions of time series patterns from diverse domains.

Unlike traditional forecasting methods that require training on your specific data, GradientCastFM understands time series patterns zero-shot. Send your historical data, specify a horizon, and get predictions immediately.

The model handles seasonality, trends, and complex patterns automatically. It works across frequencies from minutes to years, and supports both single and multi-series forecasting.

Technical Highlights

  • Transformer Architecture
    500M+ parameters trained on diverse time series
  • GPU-Accelerated
    Sub-second inference on NVIDIA A100
  • Covariate Support
    Static and dynamic external variables

Features

Multi-Series Forecasting

Forecast multiple time series in a single API call. The model captures cross-series patterns and correlations automatically.

Covariate Support

Improve accuracy with external variables. Support for static (e.g., store size) and dynamic (e.g., temperature, promotions) covariates.

Flexible Frequencies

Works with any frequency: minute, hourly, daily, weekly, monthly, quarterly, or yearly. Automatic pattern detection for each.

Batch Processing

Process hundreds of series in a single request. Optimized for production pipelines with efficient batching.

Pandas Integration

Native DataFrame support. Pass DataFrames directly and receive forecasts as DataFrames. Zero conversion overhead.

Auto Context Length

The model automatically determines optimal context. Just provide your data and the model handles the rest.

Code Examples

forecast.py
from gradientcast import GradientCastFM

fm = GradientCastFM(api_key="your-api-key")

result = fm.forecast(
    input_data={"daily_sales": [100, 120, 115, 130, 125, 140]},
    horizon_len=7,
    freq="D"
)

print(result["daily_sales"])
# [145.2, 148.7, 151.3, 154.8, 157.2, 160.1, 162.5]

Model One-Off Events with Covariates

Real-world forecasting often involves events that break normal patterns—promotions, holidays, product launches, or unexpected disruptions. GradientCastFM handles these through covariates.

Static covariates represent fixed attributes like store size or product category. Dynamic covariates capture time-varying factors like temperature, marketing spend, or binary event flags.

Simply add a binary indicator (0 or 1) for special events in your dynamic covariates, and the model learns to adjust predictions accordingly—no manual adjustment or rule-writing needed.

Covariate Types

  • Static Numerical
    Store size, capacity, baseline metrics
  • Dynamic Numerical
    Temperature, price, marketing spend
  • Event Indicators
    Holidays, promotions, outages (0/1 flags)

Best For

  • Demand forecasting and inventory planning
  • Revenue and sales projections
  • User growth and engagement predictions
  • Resource capacity planning
  • Financial time series analysis
  • Any domain with historical patterns

Technical Specifications

<1s
p90 Latency
Single series
100 series
Max Batch Size
Per request
2048 points
Max Context
Auto-optimized
180 seconds
Timeout
Configurable

Ready to start forecasting?

Get your API key and make your first forecast in minutes.

Get API KeyView Pricing