1
$\begingroup$

I am building a time series forecasting model in which I am considering the macroeconomic indicators as predictors.I wanted to ask 2 things

  1. How do I get the future values?I have seen trading economics and some other sites but they are paid. Also if I want predict indicators on my own how should I go about that?

  2. All the data for indicators are given in two forms on govt websites - seasonally adjusted and seasonally not adjusted. I am using seasonally adjusted data for my model building.Is that correct?

3.I am building a product demand forecasting model. I have chosen 20-30 macroeconomic indicators and other exogenous variables. I wanted to understand how do I check with explanatory variables will help me in forecasting the demand. How do I compare two time series .Can I build a separate regression models for each indicator with dependent variable as demand and then select variables which are significant into my final forecasting model. Or can I use Cross Correlation or Granger's causality approach to determine significant predictors?

Any help would be appreciated

$\endgroup$
1
$\begingroup$

I will give a partial answer.

  1. I assume you mean “forecast values.” Forecasts generally come from private firms that do the forecasts, and are typically expensive. There are some surveys that are in the public domain (Philadelphia Fed Survey of Professional Forecasters). You can generate your own using a model, but quality depends uponyour model.
  2. It is normally easier to work with seasonally adjusted data. However, the data are passed through a filter, and distortions can occur. (For example, the pandemic of 2020 has caused spikes in data, and seasonal adjustment can make those spikes worse.)
  3. The last question is open-ended. People spend their entire careers in finance and at central banks building models.
| improve this answer | |
$\endgroup$
  • $\begingroup$ Thanks for your reply. I am really sorry that I wasn't clear with my 3rd question. I've edited it now. Could you please have a look? $\endgroup$ – Raj Jul 12 at 16:50
  • $\begingroup$ My suggestion is to break out that as a new question. Check that no similar questions exist, then describe the problem. If the title of the question reflects the question, more people will look at it. As for myself, I’d probably just wing it, which is not particularly rigorous... $\endgroup$ – Brian Romanchuk Jul 12 at 20:18
  • $\begingroup$ Thanks.. Will do that $\endgroup$ – Raj Jul 13 at 8:09
0
$\begingroup$
  1. If you are forecasting some special data about certain country, i guess, you can take some official forecasts made by government agencies or IMF. In case of IMF there are usually only forecast of GDP growth, while official agencies may post forecasted values for wider range of ts. Also if you have access for Bloomberg terminal in your university or work, then you can watch a forecasts made by different private entities. These forecasts both made by officials and private structures usually have annual frequency.

  2. If you use non adjusted data, then you will ger spurious correlations between factors and hence bad forecast with unrelated variables in model.

  3. I would choose variables in the model based on macroeconomic theory. For example, if you forecast nominal GDP, and you see great correlation between nominal GDP and M2, then you should understand that they are correlated because they are both affected by inflation and also because they are non-stationary (they have non-stochastic trend, variables also can be non-stationary in case of absence of non-stochastic trend and in this case correlation between them also can be spurious. You should use stationarity tests to see if variables are non-stationary, for example, ADF-test. Choose version of test with constant, include linear trend in test to see, whenever time series is linear-trend stationary. If it is, then you can detrend time series). Usual approach to solve non-stationarity problems is to make your regression in differences. Also you can use VAR approach. If you use VAR without exogenous variables, then you need no assumptions about future paths of independent time series.

| improve this answer | |
$\endgroup$
  • $\begingroup$ Hi Maksim , first of all thanks for your comprehensive reply. Just to reiterate if I have understood correctly - First and foremost thing is making the predictors and the target stationary through differencing/transformation and then check using ADF, KPSS tests to so that we do not get spurious correlation. Then I can run regression for each one of the combination of target ad predictors and select the significant predictors. Also, Can you have a look at the 3rd question again if possible? $\endgroup$ – Raj Jul 12 at 17:02
  • $\begingroup$ 1. Firstly, you need to check whenever your variables are already stationary. Usually macroeconomic variables are unstationary in their original form. So it is highly likely that their first difference is stationary. Then you should need to run regression on all independent variables because if you have ommited variables in regression then the coefficients of regressors tend to be biased if ommited variable is correlated with included dependent. When you built model with all coefficients then you can leave only significant ones. And in the end you need to be aware of endogenity. $\endgroup$ – Maksim Tarasenko Jul 12 at 21:12
  • $\begingroup$ If you are trying to identificate demand, then coefficients in your regression can represent not demand but equilibrium of output, which depends both on demand and supply. Check this rstudio-pubs-static.s3.amazonaws.com/…. $\endgroup$ – Maksim Tarasenko Jul 12 at 21:16
  • $\begingroup$ "With included dependent". I meant independent. You can use cross-correlation to determine which variables are good explanators to include them in the model simultaneously by the way. $\endgroup$ – Maksim Tarasenko Jul 12 at 21:19

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.