forecasting-models

name	forecasting-models
description	Build time-series forecasting models for business metrics. Outputs model selection framework, Prophet/ARIMA implementation, evaluation methodology, production serving pattern, and uncertainty quantification.
argument-hint	["metric to forecast","data frequency","horizon","seasonality patterns","accuracy requirements"]
allowed-tools	Read, Write, Bash

Forecasting Models

Forecasting predicts future values of a time series: revenue, demand, traffic, capacity. Good forecasting requires understanding trend, seasonality, holidays, and exogenous variables. The best model is often not the most complex — a well-tuned baseline beats a poorly-tuned neural network.

Model Selection Guide

< 2 years data, clear seasonality, need interpretability
→ Prophet (Meta) — excellent defaults; handles missing data; holidays

Stationary data, no clear seasonality
→ ARIMA / SARIMA — classical; well-understood; interpretable

Multiple related series (1000 products)
→ LightGBM/XGBoost with lag features — scales; captures cross-series

Long sequences, complex patterns, large data
→ Temporal Fusion Transformer — highest accuracy; needs much more data

ALWAYS BEAT THESE BASELINES FIRST:
  Naive:           forecast = last observed value
  Seasonal naive:  forecast = same period last year
  Moving average:  forecast = mean of last N periods

Prophet Implementation

from prophet import Prophet
from prophet.diagnostics  cross_validation, performance_metrics
 pandas  pd
 numpy  np

 () -> :
    
    model = Prophet(
        growth=,
        seasonality_mode=,  
        yearly_seasonality=,
        weekly_seasonality=,
        daily_seasonality=,
        changepoint_prior_scale=,   
        interval_width=,            
    )

    model.add_seasonality(name=, period=, fourier_order=)
    model.add_country_holidays(country_name=)

    
    model.add_regressor()
    model.add_regressor()

    model.fit(df)

    
    future = model.make_future_dataframe(periods=)
    future[] = df[].mean()
    future[] = 

    forecast = model.predict(future)

    
    cv_results = cross_validation(
        model,
        initial=,  
        period=,     
        horizon=,    
        parallel=,
    )

    metrics = performance_metrics(cv_results)
     {
        : model,
        : forecast[[, , , ]],
        : metrics[].mean(),
        : metrics[].mean(),
    }