Show HN: Farseer, rust rewrite of prophet with some extra goodies

16 hours ago 2

High-performance time series forecasting with Prophet-like capabilities.
Built in Rust for maximum speed, with a familiar Python API.

Python 3.11+ Rust-powered

Why Farseer?

🚀

Blazing Fast

Rust-powered performance with automatic multithreading. Get forecasts in seconds, 5-10x faster than Prophet.

🎯

Accurate & Robust

Bayesian approach with uncertainty intervals. Handles missing data, outliers, and trend shifts automatically.

⚖️

Weighted Observations

Native support for observation weights. Emphasize recent data or downweight outliers with ease.

📊

Polars & Pandas

Works with both Polars (recommended, 5-10x faster) and Pandas DataFrames for backward compatibility.

🔧

Fully Tunable

Multiple trend types, custom seasonality, holiday effects, and regressors. Add your domain knowledge.

🔄

Prophet Compatible

Nearly identical API to Prophet. Migrate your existing code with minimal changes.

📅

Conditional Seasonality

Apply seasonal patterns only when conditions are met (e.g., weekday vs weekend patterns).

🎄

Flexible Holidays

Independent holiday priors with customizable scales. Different effects for different events.

📈

Smart Regressors

Auto-detects binary vs continuous regressors. Standardizes appropriately for optimal results.

📏

Floor & Cap

Logistic growth with both upper (cap) and lower (floor) bounds for saturating forecasts.

💾

Model Serialization

Save and load trained models as JSON for deployment and reproducibility.

⚡

Parallel Optimization

Automatically uses all CPU cores for faster model fitting. Scales with your hardware.

Farseer vs Prophet

Feature Farseer Prophet
Performance Rust-powered, 5-10x faster Python/Stan
Multithreading Automatic parallel optimization Single-threaded by default
Weighted Data Native support Not directly supported
DataFrames Polars + Pandas Pandas only
Conditional Seasonality Fully supported Fully supported
Floor Parameter Full support (logistic growth) Full support
Regressor Standardization Auto-detects binary/continuous Manual configuration
Holiday Priors Independent per-holiday scales Independent per-holiday scales
Deployment Minimal dependencies Requires Stan, PyStan
API Scikit-learn-like, Prophet-compatible Scikit-learn-like

Quick Start

Installation

# From PyPI pip install farseer # Development install from source git clone https://github.com/ryanbieber/seer cd seer export PYO3_USE_ABI3_FORWARD_COMPATIBILITY=1 # For Python 3.13+ maturin develop --release

Basic Usage (Polars - Recommended)

from farseer import Farseer import polars as pl from datetime import datetime # Create data df = pl.DataFrame({ 'ds': pl.date_range( datetime(2020, 1, 1), periods=100, interval='1d', eager=True ), 'y': range(100) }) # Fit and forecast m = Farseer() m.fit(df) future = m.make_future_dataframe(periods=30) forecast = m.predict(future) print(forecast.select(['ds', 'yhat', 'yhat_lower', 'yhat_upper']).tail())

Prophet-Compatible (Pandas)

from farseer import Farseer import pandas as pd # Create data df = pd.DataFrame({ 'ds': pd.date_range('2020-01-01', periods=100), 'y': range(100) }) # Fit and forecast - same API! m = Farseer() m.fit(df) future = m.make_future_dataframe(periods=30) forecast = m.predict(future) print(forecast[['ds', 'yhat', 'yhat_lower', 'yhat_upper']].tail())

Examples

🎯 Weighted Observations

Give more importance to recent or reliable data

import polars as pl import numpy as np from farseer import Farseer df = pl.DataFrame({ 'ds': pl.date_range(...), 'y': np.random.randn(100).cumsum() + 50, 'weight': [2.0 if i < 50 else 1.0 for i in range(100)] }) # Weights automatically detected! m = Farseer() m.fit(df)

Learn more →

� Conditional Seasonality

Different patterns for different conditions

from farseer import Farseer m = Farseer() # Add conditional seasonality m.add_seasonality( name='weekly_on_weekday', period=7, fourier_order=3, condition_name='is_weekday' ) # Add condition column df['is_weekday'] = (df['ds'].dt.weekday < 5) m.fit(df)

Learn more →

�📈 Custom Regressors

Add additional variables to improve forecasts

from farseer import Farseer m = Farseer() # Auto-detects binary vs continuous m.add_regressor('temperature', standardize='auto') m.add_regressor('is_weekend', standardize='auto') m.add_regressor('promo', prior_scale=5.0) m.fit(train_df) forecast = m.predict(test_df)

Learn more →

� Floor & Cap

Saturating growth with upper and lower bounds

from farseer import Farseer # Logistic growth with floor m = Farseer(growth='logistic') df['floor'] = 1.5 # Minimum value df['cap'] = 10.0 # Maximum value m.fit(df) future['floor'] = 1.5 future['cap'] = 10.0 forecast = m.predict(future)

Learn more →

�🔄 Manual Changepoints

Specify known trend changes in your data

from farseer import Farseer # Specify exact changepoint dates changepoints = ['2021-01-01', '2022-01-01'] m = Farseer( changepoints=changepoints, yearly_seasonality=True ) m.fit(df) forecast = m.predict(future)

Learn more →

🎄 Holiday Effects

Model special events with independent priors

from farseer import Farseer m = Farseer() # Different priors for different holidays m.add_holidays( 'christmas', dates=['2020-12-25', '2021-12-25'], prior_scale=20.0 # Strong effect ) m.add_holidays( 'minor_event', dates=['2020-03-17'], prior_scale=5.0 # Weak effect ) m.fit(df)

Learn more →

Ready to get started?

Start forecasting with the speed and reliability of Rust

Read Entire Article
  1. Homepage
  2. Technology
  3. Show HN: Farseer, rust rewrite of prophet with some extra goodies

Related

Jaho Coffee Roaster

Jaho Coffee Roaster

14 minutes ago 0
Show HN: Game where players write real JavaScript to battle other players online

Show HN: Game where players write real JavaScript to battle ...

15 minutes ago 0
Commodore LCD Emulator

Commodore LCD Emulator

17 minutes ago 0
Hostinger Web Hosting