# forecast results does not change as the covariate variables change

I am trying to forecast a time series using the bsts package. The data is as follows:

The data is collected monthly and the goal is to forecast y by v1, v2, v3, v4, v6 and v7. y is a count variable, so I use poisson family.

library(bsts)
library(dplyr)
data_f<-structure(list(v1 = c(17242647.0588235, 18498333.3333333, 21890000,
27110416.6666667, 34084800, 39341904.7619048, 45212500, 35353684.2105263,
35353684.2105263, 35811304.3478261, 40493913.0434783, 44562500,
46071176.4705882, 48077600, 45101904.7619048, 42889230.7692308,
40886400, 40281250, 39787916.6666667, 39927391.3043478, 43662800,
47327307.6923077, 50170833.3333333, 58340476.1904762, 62048421.0526316,
66089583.3333333, 72500000, 92497307.6923077, 104050400, 112614000,
137931739.130435, 125291304.347826, 113013076.923077, 111041880,
108333800, 105900409.090909, 106789000, 99218400, 105324800,
106098000, 113881000, 119995000), v2 = c(5.5e+07, 5.9e+07, 6.5e+07,
6.9e+07, 7.4e+07, 8.1e+07, 8.6e+07, 9.2e+07, 9.5e+07, 9.8e+07,
9.9e+07, 1.09e+08, 1.13e+08, 1.27e+08, 1.33e+08, 1.33e+08, 1.3e+08,
1.26e+08, 1.27e+08, 1.25e+08, 1.35e+08, 1.39e+08, 1.43e+08, 1.57e+08,
1.53e+08, 1.67e+08, 1.91e+08, 2.01e+08, 2.31e+08, 2.43e+08, 2.67e+08,
2.72e+08, 2.69e+08, 2.74e+08, 2.84e+08, 3.02e+08, 2.93e+08, 2.87e+08,
2.96e+08, 3.05e+08, 3.09e+08, 3.17e+08), v3 = c(1582, 1582, 1582,
1672, 1672, 1672, 1764, 1764, 1764, 1882, 1882, 1882, 1979, 1979,
1979, 2126, 2126, 2126, 2156, 2211, 2262, 2472, 2472, 2472, 3467,
3467, 3467, 3467, 3467, 3467, 3467, 3467, 3467, 3467, 3467, 3467,
3489, 3722, 3835, 3956, 3985, 4001), v4 = c(51663, 60976, 66582,
81250, 102708, 120493, 121996, 123500, 111061, 108454, 118300,
130323, 134135, 144001, 133362, 125505, 118750, 113765, 113928,
114579, 126722, 130753, 135349, 150653, 155444, 160430, 174200,
209365, 215418, 237377, 292610, 268402, 255041, 250225, 238348,
243862, 243910, 221980, 234970, 247850, 259710, 274700), v6 = c(0.081,
0.08, 0.082, 0.087, 0.097, 0.113, 0.134, 0.156, 0.18, 0.206,
0.235, 0.269, 0.306, 0.342, 0.376, 0.404, 0.422, 0.427, 0.42,
0.411, 0.4, 0.386, 0.37, 0.348, 0.322, 0.298, 0.278, 0.264, 0.258,
0.26, 0.272, 0.29, 0.305, 0.322, 0.342, 0.364, 0.495, 0.469,
0.476, 0.428, 0.432, 0.437), v7 = c(10000, 10000, 10000, 10000,
10000, 10000, 10000, 10000, 10000, 10000, 10000, 10000, 10000,
10000, 10000, 10000, 10000, 10000, 10000, 10000, 30000, 30000,
30000, 30000, 30000, 30000, 30000, 30000, 30000, 30000, 30000,
30000, 30000, 30000, 30000, 30000, 30000, 30000, 30000, 30000,
30000, 30000), y = c(10, 20, 39, 21, 18, 10, 30, 28, 19, 15,
16, 19, 7, 15, 10, 14, 11, 11, 14, 52, 8, 10, 8, 11, 7, 14, 11,
15, 14, 15, 13, 12, 13, 17, 22, 18, 0, 0, 0, 0, 0, 0)), row.names = c(NA,
42L), class = "data.frame")
data_f_train<-data_f[1:36,]
data_f_test<-data_f[37:42,]
data_f_test<-data_f_test %>% select(-y)
ss=AddLocalLinearTrend(list(),y=log1p(data_f_train$$y)) ss=AddSeasonal(ss,log1p(data_f_train$$y),nseasons = 12, season.duration = 1)
model2=bsts(y~.
,state.specification =ss,niter=100,family="poisson",data=data_f_train)

pred <- predict(model2,  newdata=data_f_test)
par(mfrow=c(2,1))
plot(pred,ylim=c(0,100))
data_f_test2<-data_f_test+10000
pred2 <- predict(model2,  newdata=data_f_test2)
plot(pred2,ylim=c(0,100))


It seems my steps is wrong completely since the forecast does not change as I change the covariate variables.

> head(model2\$coefficients)
(Intercept) v1 v2 v3 v4 v6 v7
[1,]           0  0  0  0  0  0  0
[2,]           0  0  0  0  0  0  0
[3,]           0  0  0  0  0  0  0
[4,]           0  0  0  0  0  0  0
[5,]           0  0  0  0  0  0  0
[6,]           0  0  0  0  0  0  0

• Is this because the y variable is small but the covariate variables are too big?! Nov 3, 2021 at 16:45

You are fitting a poisson model for count data, which means the predicted value can't be $$< 0$$. If there is a negative relationship between a covariate and the outcome, and the predicted count is already $$0$$, increasing the value of the covariate won't change the predicted value, since it can't go any lower.
To test this, you could try fitting the same data with a Gaussian linear model, probably by setting family = 'gaussian', or similar. This allows negative predicted values, and so the predictions should no longer be constant for these cases.
• Hi Masoud. I've just looked at the actual code you've posted. The y column, which you've annotated as "forecast" is actually the data, not the model predictions, so I'm afraid I don't understand your question, sorry.