使用forecast包nnetar函数时预测区间过窄的问题求助

Hey there! Let's tackle that narrow prediction interval issue with your nnetar model for weekly e-commerce revenue. I’ve worked through similar time series challenges before, so here’s a breakdown of what’s going on and how to fix it:

First, a quick recap: By default, nnetar calculates prediction intervals using the variance of the model’s fitting residuals, plus some accounting for neural network prediction uncertainty. But there are a few key reasons this might end up too tight for your e-commerce data:

• The default number of neural network repeats (repeats=20) might not capture enough of the randomness in neural network weight initialization, leading to underrepresented prediction variance.
• If your explanatory variables (conversion rate, ad spend) have their own future uncertainty, nnetar doesn’t automatically incorporate that into the interval—so it’s only accounting for model error, not Xreg volatility.
• Neural networks can sometimes "overfit" to the training data, making residuals look smaller than they should be, which shrinks the interval.

Let’s go through actionable steps, starting with the easiest tweaks:

1. Increase the Number of Neural Network Repeats #

The repeats parameter controls how many separate neural networks are trained (each with random initial weights). More repeats mean the model captures more of the inherent randomness in neural network predictions, which naturally widens the interval. Try bumping it up to 50 or 100:

# Full example with increased repeats
fit_test <- nnetar(total_revenue_ts, size = 5, repeats = 50, xreg = cbind(conversion_rate, ad_spend))
# Generate forecast with your future Xreg values
fc <- forecast(fit_test, xreg = future_xreg_data, h = 12) # h = number of weeks to predict

2. Use Bootstrap Prediction Intervals #

Instead of relying on residual variance alone, enable bootstrap intervals in the forecast() function. This resamples residuals to simulate real-world prediction uncertainty, which often produces more realistic intervals for nonlinear models like neural networks:

fc <- forecast(fit_test, xreg = future_xreg_data, h = 12, bootstrap = TRUE, npaths = 500)

The npaths parameter sets how many bootstrap samples to generate—more samples mean more stable intervals (but take a bit longer to run).

3. Account for Xreg Uncertainty #

If your future conversion rate and ad spend values are predictions themselves (not fixed budgets/forecasts), their own uncertainty should be included in your revenue prediction intervals. Here’s a quick way to do this with scenario simulation:

# Step 1: Generate multiple future scenarios for your Xreg variables
n_scenarios <- 100
# Predict future conversion rates with uncertainty
future_conversion <- forecast(auto.arima(conversion_rate_ts), h=12)$simulate(n_scenarios)
# Predict future ad spend with uncertainty
future_ad_spend <- forecast(auto.arima(ad_spend_ts), h=12)$simulate(n_scenarios)

# Step 2: Generate revenue predictions for each scenario
all_forecasts <- lapply(1:n_scenarios, function(i) {
  current_xreg <- cbind(future_conversion[,i], future_ad_spend[,i])
  forecast(fit_test, xreg = current_xreg, h=12)$mean
})

# Step 3: Calculate quantiles to get your interval bounds
combined_fcs <- ts(do.call(cbind, all_forecasts), start = start(future_conversion))
lower_95 <- apply(combined_fcs, 1, quantile, 0.025)
upper_95 <- apply(combined_fcs, 1, quantile, 0.975)
mean_forecast <- apply(combined_fcs, 1, mean)

This gives you intervals that reflect both model uncertainty and the volatility of your key drivers.

4. Tweak the Confidence Level (Quick Fix) #

If you just need a wider interval without changing the model, adjust the level parameter in forecast() to include higher confidence levels. For example:

fc <- forecast(fit_test, xreg = future_xreg_data, h = 12, level = c(80, 90, 99))

This will give you 80%, 90%, and 99% intervals—though note this doesn’t fix the root cause of underrepresented uncertainty, it just adjusts how much of the existing variance you’re showing.

Before adjusting intervals, make sure your base model is solid:

• Check residuals: Run checkresiduals(fit_test) to confirm residuals are white noise (no autocorrelation, constant variance). If not, tweak the size parameter (try 3-10) or let nnetar auto-select lag terms with the p parameter.
• Standardize Xreg: Neural networks are sensitive to scale. Standardize your conversion rate and ad spend with scale() before fitting the model to improve performance and residual stability.

These steps should help you get prediction intervals that better match the actual volatility of your e-commerce revenue. Start with increasing repeats and enabling bootstrap—those are the lowest-effort, highest-impact changes!

内容的提问来源于stack exchange，提问作者Martin Brummerstedt