Off by one error in seasonal regression models?

[steve-the-bayesian]

I am a forecasting practitioner who has been trying to use your bsts package to solve some forecasting tasks. The data I play around to learn how to use bsts is the M5 data (daily sales in 1969 days) and bizarrely when I include a dummy variable to model the Christmas effect, the model suddenly fails to capture the weekly seasonal pattern. Specifically, the data is like this

read data from csv

unit_sales <- readr::read_csv('M5example.csv', col_types = cols()) unit_sales$date <- as.Date.character(unit_sales$date) image.png image.png

This is how I split the dataset

h <- 56 len_his <- dim(unit_sales)[1] - h data_train <- head(unit_sales, len_his) data_test <- tail(unit_sales, h) y_fit <- zoo(data_train$sales, order.by = data_train$date) y_true <- zoo(data_test$sales, order.by = data_test$date)

This is how I build the first model

ss <- AddSemilocalLinearTrend(list(), y_fit)

ss <- AddSeasonal(ss, y_fit, nseasons = 7) model <- bsts(y_fit, state.specification = ss, niter = 1000, seed=100) y_pred <- predict(model, horizon=h)

The model correctly capture the weekly seasonal pattern

image.png

I then tried to build a second model that includes the Christmas regressor

ss2 <- AddSemilocalLinearTrend(list(), data_train$sales) ss2 <- AddSeasonal(ss2, data_train$sales, nseasons = 7) model2 <- bsts(sales~., state.specification = ss2, data = data_train[, -c(1)], niter = 1000, seed = 100) y_pred2 <- predict(model2, newdata = data_test[, -c(1)] ,horizon=h)

It seems that there is a one-day shift between the true seasonal pattern and the forecasted one.

image.png

Please see attached the csv for data and the ipynb file for reproducible code.

I find it difficult to understand why including one regressor makes the model fail to capture the seasonal pattern. Is there anything I do wrong when building the second model? If not, what is the reason to cause the model to behave like this?

Thanks a lot for your time. I am looking forward to your reply. M5example.csv

[steve-the-bayesian]

The version I used was 0.9.2. I've updated it to 0.9.6 and suddenly the problem is gone. Now the seasonal pattern is correctly forecasted.