I'm a bit confused on how to properly compare a known trend to an estimated one. I've got two sets of data which are of the following format,

$$y_{t_1}=x_{t_1}+WN$$ $$y_{t_2}=x_{t_1}+MA(2)$$

Where $WN \sim N(0,1)$, $MA(1)$ is a simulated 2nd order moving average and $x_{t_1}$ is a known trend function. Now I have gone ahead and estimated the trend function $\hat{x_{t_1}}$ using Splines via generalized cross validaiton (GCV). So my two data setups are,

$$y_{t_1}=\hat{x_{t_1}}+WN$$ $$y_{t_2}=\hat{x_{t_2}}+MA(2)$$

Where $\hat{x_{t_1}}$ is the modeled trend function by Splines. I now want to compare the modeled trend, $\hat{x_{t_1}}$,$\hat{x_{t_2}}$, to the actual known trend, $x_{t_1}$ and come to the conclusion that

A few things I have already tried,

  1. Fixed $x_{t_1}$ as an increasing trend.
  2. Added WN and MA(2) to create $y_{t_1}$, $y_{t_2}$
  3. Modelled $x_{t_1}$, without the error terms, via Splines to get $\hat{x_{t_1}}$.
  4. I then added $\hat{x_{t_1}}$ to the WN and MA(2) separately and calculated the MSE.
  5. However the MSE values ended up being the same when compared to the original values, which is what I was not expecting as this concludes there is no difference in the trends.

Note: For #3 I also tried creating two different models, $m_1$ and $m_2$, with the errors included. I then subtracted the error terms from both models to get two estimated trends, which were very similar to one another. I then planned on comparing these two trends to the original fixed one however after putting some more thought into it, this approach doesn't really make sense.

Question: How can I compare the two trends, $\hat{x_{t_1}}$,$\hat{x_{t_2}}$, and $x_{t_1}$ to either prove or disprove my assumption of that $\hat{x_{t_1}}$,$\hat{x_{t_2}}$ is a good trend estimate for the setup with White Noise and not a good trend estimate for the setup with an MA(2) process?

Many thanks in advance.


1 Answer 1


After mangling around with this for a short while, I finally managed to get everything to work properly. Here were the steps I took,

  1. Fixed $x_{t_1}$ as an increasing trend.
  2. Added WN and MA(2) to create $y_{t_1}$, $y_{t_2}$
  3. Modelled $x_{t_1}$ with both of the error terms separately, via Splines to get $\hat{x_{t_1}}$ and $\hat{x_{t_2}}$.
  4. Plotted $x_{t_1}$ with $\hat{x_{t_1}}$ on one plot and $x_{t_1}$ with $\hat{x_{t_2}}$ on another plot.
  5. The differences were quite hard to see graphically so I decided to calculate the MSE of both plots.
  6. Compared the two MSEs with one another to see which one performed better.

The result of 3 different simulations was that the MSE of $\hat{x_{t_1}}$ estimated tend was lower than the estimated trend $\hat{x_{t_2}}$ when compared to the actual known trend of $x_{t_1}$. Which confirms the original assumption that estimating a trend via WN error terms is better than estimating a trend with MA(2) error terms. The details are quite vague but for the purpose of the work I was doing it is enough to make this conclusion. More details can always be added if needed or requested.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.