Time Series Forecasting: Beyond Traditional ARIMA

Comparing modern deep forecasting models and gradient boosted trees against classical autoregressive techniques.

Author: Somesh Chatterjee
Tags: time-series, forecasting, data-science, statistics

Time Series Forecasting: Beyond Traditional ARIMA

Accurate forecasting across fluctuating demand cycles requires modeling non-linear trends, calendar effects, and complex multivariate relationships that classic ARIMA models often struggle to capture.

When ARIMA Falls Short

Classical ARIMA relies on strict stationarity assumptions and linear autoregression:

  • Fixed lag structures cannot easily incorporate high-cardinality exogenous features.
  • Multi-seasonal periodicity (hourly, weekly, annual) requires exponential parameter scaling.
  • Cross-series correlations between thousands of items cannot be shared.

The Rise of Gradient Boosted Forecasting

Tree-based models (like LightGBM and XGBoost) have proven consistently dominant in real-world forecasting benchmarks when fed with engineered temporal features:

  • Lag features: Values at t-1, t-7, t-30.
  • Rolling statistics: Rolling mean, standard deviation, and exponential moving averages.
  • Calendar encodings: Sinusoidal cyclical representations of day-of-week and month.
import numpy as np
import pandas as pd

def add_cyclical_features(df, col, max_val):
    df[f"{col}_sin"] = np.sin(2 * np.pi * df[col] / max_val)
    df[f"{col}_cos"] = np.cos(2 * np.pi * df[col] / max_val)
    return df

Summary

Modern time series modeling combines robust feature engineering with ensemble techniques, offering greater resilience to structural shifts than rigid parametric methods alone.