tslm中协变量失效求助：趋势季节模型完美拟合致协变量无影响

Hey there, let's work through this tricky issue you're hitting with the tslm function from Hyndman's forecast package. It’s super frustrating when variables you know should matter don’t show up as significant, and even weirder when removing them doesn’t shift your predictions at all. Let’s break this down step by step.

First off, hitting an R²=1 with just trend + season means those two components already perfectly explain every single sales observation in your 2014-2016 training set. That leaves no room for other covariates like items or delivery volume to add any extra explanatory power—your model literally can’t see their independent effect because trend and season already capture everything in the training data.

• Redundant Information / Multicollinearity: Your covariates (like items) might be highly correlated with trend or season. For example, maybe you stock more items every holiday season (so season already captures that) or item counts grow linearly year over year (trend covers that). When variables overlap this much, the model treats the covariate as unnecessary.
• Overfitting with Weekly Seasonal Dummies: You set freq=365.25/7 (~52.178), which makes tslm generate 52 seasonal dummy variables for weekly patterns. With 3 years of data (156 observations), 52 dummies plus a trend term gives you enough variables to perfectly fit the training data—this is overfitting in action, and it swamps any signal from your covariates.
• Mismatched Time Scales or Weak Signals: If your covariates change way less than the seasonal/trend fluctuations, or if you aggregated them incorrectly (e.g., daily item counts crammed into weekly averages), their impact gets lost in the noise of the stronger trend/season signals.

Let’s walk through actionable steps to uncover your covariates’ real impact:

1. Replace Seasonal Dummies with Fourier Terms
   Instead of using 52 dummy variables (which overfit), use Fourier series to capture seasonal patterns with far fewer variables. This frees up the model to notice your covariates.

   # Use K=3 (adjust K up/down to balance fit vs overfitting)
   m.lm <- tslm(ts.train ~ trend + fourier(ts.train, K=3) + items, data=df.train)
   

   K controls how detailed the seasonal fit is—start with K=2-4 and tweak based on model diagnostics.

2. Remove Trend/Season from Your Covariates
   If your covariates are correlated with trend/season, extract the part of them that isn’t explained by those components, then use that residual in your model:

   # Convert items to a time series and decompose it
   ts.items <- ts(df.train$items, freq=52)
   items_decomp <- stl(ts.items, s.window="periodic")
   # Grab the residual (items variation not explained by trend/season)
   items_resid <- items_decomp$time.series[, "remainder"]
   # Use the residual in your model
   m.lm <- tslm(ts.train ~ trend + season + items_resid, data=df.train)
   

   This isolates the unique impact of items on sales, separate from seasonal or trend-driven changes.

3. Try a More Flexible Model (Like ARIMA with Covariates)
   Linear models struggle when trend/season dominate the data. Instead, use auto.arima which handles seasonality more gracefully and can incorporate covariates:

   m_arima <- auto.arima(ts.train, xreg=df.train$items, seasonal=TRUE)
   p_arima <- forecast(m_arima, h=48, xreg=df.test$items)
   

   ARIMA models are designed for time series data and won’t get as easily swamped by perfect trend/season fits.

4. Double-Check Your Training Data
   An R²=1 in real-world retail data is extremely rare—are you sure your training data hasn’t been pre-processed (e.g., adjusted for trend/season already)? Or is there a quirk in how the data was collected? Verify that individual weekly sales values aren’t exactly matching a seasonal+trend formula.

When removing items doesn’t change your forecast, it’s because the model assigned a coefficient of (near) zero to items. That’s a clear sign the variable adds nothing new once trend and season are included—so fixing the overfitting or redundancy issues above is key to unlocking its impact.

内容的提问来源于stack exchange，提问作者Tim Brennan