gradientcast 0.1.3

Official Python SDK for GradientCast ML Services - Time series forecasting and anomaly detection

GradientCast Python SDK

Official Python SDK for GradientCast AI Services - 0-shot time series forecasting and anomaly detection.

Features

• GradientCastFM - Enterprise-level 0-shot time series forecasting using state-of-the-art foundation models
• GradientCastPulseAD - Flagship plug-and-play anomaly detection
• GradientCastDenseAD - Intelligent pattern anomaly detection with severity classification

Installation

pip install gradientcast

With pandas support:

pip install gradientcast[pandas]

Quick Start

Forecasting (GradientCastFM)

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.forecast["daily_sales"])
# [145.2, 148.7, 151.3, ...]

print(f"Processing time: {result.model_info.processing_time}s")

Multi-series Forecasting

result = fm.forecast(
    input_data={
        "product_a": [100, 120, 115, 130],
        "product_b": [200, 220, 215, 230]
    },
    horizon_len=7,
    freq="D"
)

# Access forecasts by series name
print(result["product_a"])
print(result["product_b"])

With Covariates

result = fm.forecast(
    input_data={"sales": [100, 120, 115, 130, 125, 140]},
    horizon_len=7,
    freq="D",
    static_numerical_covariates={
        "store_size": {"sales": 5000.0}
    },
    dynamic_numerical_covariates={
        "temperature": {"sales": [72, 75, 78, 80, 82, 79, 76, 74, 77, 79, 81, 78, 75]}
    }
)

Dense Anomaly Detection (GradientCastDenseAD)

Historical data serves as context for pattern learning. Anomaly detection is performed only on the most recent data point(s) in the time series.

from gradientcast import GradientCastDenseAD

ad = GradientCastDenseAD(api_key="your-api-key")

# Earlier points provide context; latest point(s) are evaluated for anomalies
result = ad.detect([
    {"timestamp": "01/01/2025, 12:00 AM", "value": 1500000},  # Context
    {"timestamp": "01/01/2025, 01:00 AM", "value": 1520000},  # Context
    {"timestamp": "01/01/2025, 02:00 AM", "value": 100000},   # <-- Evaluated for anomaly
    # ... more data points
])

if result.has_anomaly:
    print(f"Alert: {result.alert_severity}")
    for point in result.anomalies:
        print(f"  {point.timestamp}: {point.value} (severity: {point.magnitude.severity})")

Tuning Detection Parameters

result = ad.detect(
    data=[...],
    contamination=0.05,        # Expected proportion of anomalies
    n_neighbors=20,            # Density calculation neighbors
    min_contiguous_anomalies=3 # Require 3+ consecutive anomalies
)

PulseAD Anomaly Detection (GradientCastPulseAD)

Historical data serves as context for expected behavior. Anomaly detection is performed only on the most recent data point(s) in the time series.

from gradientcast import GradientCastPulseAD, ThresholdConfig

ad = GradientCastPulseAD(api_key="your-api-key")

# Earlier points provide context; latest point(s) are evaluated for anomalies
result = ad.detect(
    time_series_data={
        "user_count": [
            {"timestamp": "01/01/2025, 12:00 AM", "value": 1500000.0},  # Context
            {"timestamp": "01/01/2025, 01:00 AM", "value": 1520000.0},  # Context
            {"timestamp": "01/01/2025, 02:00 AM", "value": 1480000.0},  # Context
            {"timestamp": "01/01/2025, 03:00 AM", "value": 1510000.0},  # Context
            {"timestamp": "01/01/2025, 04:00 AM", "value": 1530000.0},  # Context
            {"timestamp": "01/01/2025, 05:00 AM", "value": 800000.0},   # <-- Evaluated
        ]
    }
)

if result.has_anomaly:
    for anomaly in result.anomalies:
        print(f"{anomaly.dimension}: {anomaly.percent_delta} deviation")
        print(f"  Actual: {anomaly.actual_value}, Predicted: {anomaly.predicted_value}")

Custom Thresholds

config = ThresholdConfig(
    default_percentage=0.20,  # 20% deviation threshold
    default_minimum=50000,
    per_dimension_overrides={
        "AllUp": {
            "percentage_threshold": 0.10,  # Stricter for AllUp
            "minimum_value_threshold": 3000000
        }
    }
)

result = ad.detect(time_series_data, threshold_config=config)

Pandas Integration

All clients support pandas DataFrames:

import pandas as pd
from gradientcast import GradientCastFM

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

# From DataFrame
df = pd.DataFrame({
    "date": pd.date_range("2024-01-01", periods=30, freq="D"),
    "sales": [100 + i * 2 for i in range(30)],
    "product": ["A"] * 15 + ["B"] * 15
})

result_df = fm.forecast_df(
    df,
    value_column="sales",
    series_column="product",
    horizon_len=7,
    freq="D"
)

print(result_df)
#     series  horizon_step  forecast
# 0        A             1     130.5
# 1        A             2     132.8
# ...

Configuration

Custom Endpoint Support

# Production (default)
fm = GradientCastFM(api_key="key", environment="production")

# Custom endpoint
fm = GradientCastFM(api_key="key", endpoint_url="https://custom.endpoint.com/score")

Timeout and Retries

fm = GradientCastFM(
    api_key="key",
    timeout=300,      # 5 minutes
    max_retries=5     # Retry up to 5 times on transient failures
)

Context Manager

with GradientCastFM(api_key="key") as fm:
    result = fm.forecast(data, horizon_len=10, freq="H")
# Session automatically closed

Error Handling

from gradientcast import GradientCastFM
from gradientcast import (
    GradientCastError,
    AuthenticationError,
    RateLimitError,
    ValidationError,
    TimeoutError,
    APIError,
)

fm = GradientCastFM(api_key="key")

try:
    result = fm.forecast(data, horizon_len=10, freq="H")
except AuthenticationError:
    print("Invalid API key")
except RateLimitError as e:
    print(f"Rate limited. Retry after {e.retry_after}s")
except ValidationError as e:
    print(f"Invalid input: {e}")
except TimeoutError:
    print("Request timed out - try increasing timeout")
except APIError as e:
    print(f"API error [{e.status_code}]: {e.message}")
except GradientCastError as e:
    print(f"Unexpected error: {e}")

Response Objects

ForecastResponse

result = fm.forecast(...)

result.forecast          # Dict[str, List[float]] - forecasts by series
result.model_info        # ModelInfo - execution metadata
result.raw               # Dict - raw API response
result.to_dataframe()    # Convert to pandas DataFrame
result["series_name"]    # Shorthand for result.forecast["series_name"]

DenseADResponse

result = dense_ad.detect(...)

result.alert_status      # "no_alert" or "incident_active"
result.alert_severity    # "none", "low", "medium", "high", "critical"
result.has_anomaly       # bool - convenience property
result.anomalies         # List[TimelinePoint] - confirmed anomalies only
result.timeline          # List[TimelinePoint] - all points
result.to_dataframe()    # Convert to pandas DataFrame

ADResponse

result = ad.detect(...)

result.results           # List[ADResult] - all detection results
result.has_anomaly       # bool - convenience property
result.anomalies         # List[ADResult] - anomalies only
result.processing_time_ms      # Total processing time
result.fm_processing_time_ms   # FM endpoint time
result.to_dataframe()    # Convert to pandas DataFrame

Supported Frequencies

Code	Description
H	Hourly
T / MIN	Minute
D	Daily
B	Business day
W	Weekly
M	Monthly
Q	Quarterly
Y	Yearly

Requirements

• Python 3.8+
• requests >= 2.25.0
• pandas >= 1.3.0 (optional, for DataFrame support)

License

MIT License - see LICENSE for details.

Support

• Documentation: https://docs.gradientcast.com
• Issues: https://github.com/GradientCast/gradientcast-sdk/issues
• Email: [e-mail]