关于R语言租车销量日/月/季度预测项目的技术问询

Hey there! Great job getting your first R-based car rental sales forecasting project off the ground—combining ARIMA with GBM and contextual features like holidays/weather is already a solid foundation. Let’s dive into actionable tweaks to boost your forecast accuracy across daily, monthly, and quarterly horizons:

1. Align & Enhance Your Dual Datasets #

You mentioned two datasets with different calculation logic spanning 10 years (daily + monthly). First, fix any inconsistencies to ensure your models have reliable inputs:

• Validate Dataset Consistency: Aggregate your daily data into monthly figures using dplyr::group_by(year, month) %>% summarise(monthly_sales = sum(daily_sales)), then compare against your existing monthly dataset with tools like compareDF::compare_df() to spot gaps (e.g., differing inclusion of canceled bookings, regional splits). Resolve these discrepancies first—garbage in = garbage out!
• Expand Contextual Features: Beyond basic holidays and weather, add high-impact temporal features:
  • Lagged sales: Use dplyr::lag(daily_sales, n = 7) for weekly seasonality, or lag(monthly_sales, n = 3) for quarterly trends.
  • Rolling statistics: Calculate 7-day moving averages with zoo::rollmean(daily_sales, k = 7, fill = NA) to smooth short-term fluctuations.
  • Holiday derivatives: Flag "pre-holiday peak" (1-3 days before a holiday) or "post-holiday slump" periods, plus categorize holidays (national vs. local, long weekend vs. single day).
  • Weather granularity: Add dummy variables for weather types (rain, snow, extreme heat) and cumulative rainy days instead of just daily precipitation.

2. Refine Your ARIMA Implementation #

Auto-ARIMA is powerful, but you can tailor it to your hierarchical forecasting needs (daily → monthly → quarterly):

• Incorporate External Regressors: Don’t limit ARIMA to just time series—pass your holiday/weather features as the xreg parameter:

  arima_model <- auto.arima(daily_ts, xreg = daily_features, seasonal = TRUE)
  

• Hierarchical Forecasting Consistency: Ensure your monthly/quarterly forecasts align with daily predictions (no conflicting numbers!). Use the fable package’s hierarchical tools:

  library(fable)
  hierarchy <- hierarchy(quarterly ~ monthly ~ daily)
  hf <- hierarchical_forecast(arima_model, hierarchy, reconcile = "mint")
  

  This ensures lower-level forecasts roll up to higher levels correctly.

3. Tune Your GBM Model for Better Performance #

GBM’s default settings rarely deliver optimal results—systematically tune hyperparameters with time-series-aware cross-validation:

• Use Time-Series Cross-Validation: Avoid random CV (it breaks temporal order!). Instead, use caret::trainControl(method = "timeslice", initialWindow = 365*8, horizon = 365) to train on 8 years of data and validate on the next year.
• Optimize Key Hyperparameters: Focus on these critical settings using caret::train():
  • shrinkage (learning rate): Start with 0.01-0.1 (smaller = more trees needed, but less overfitting)
  • interaction.depth: Control tree complexity (2-5 works well for most time-series tasks)
  • n.trees: Use gbm::gbm.perf() to find the optimal number before overfitting kicks in.

4. Ensemble Models for Robustness #

ARIMA excels at capturing linear trends/seasonality, while GBM handles nonlinear feature interactions—combining them often outperforms either model alone:

• Weighted Average: Calculate weights based on each model’s validation accuracy (e.g., if GBM has a 10% lower MAPE than ARIMA, assign GBM a 60% weight, ARIMA 40%):

  combined_forecast <- 0.4*arima_forecast + 0.6*gbm_forecast
  

• Meta-Model: Train a simple linear regression or glmnet model that takes ARIMA and GBM forecasts as input features, and predicts actual sales. This lets the data decide the optimal weighting.

5. Validate & Iterate with Time-Series Metrics #

Don’t rely on just RMSE—use metrics tailored to forecasting:

• Use forecast::accuracy() to compute MAPE (great for percentage-based error tracking) and MASE (scales error to historical volatility, perfect for comparing daily/monthly/quarterly forecasts).
• Backtest across multiple time windows (e.g., train on 2013-2020, validate 2021, test 2022; then train 2014-2021, validate 2022) to ensure your model isn’t overfitting to a single period.
• Detect structural breaks in your 10-year data with changepoint::cpt.meanvar(daily_sales)—if there’s a clear shift (e.g., post-2020 pandemic), split your data and build separate models for each period.

内容的提问来源于stack exchange，提问作者Emre Aktürk