0
$\begingroup$

So in recent months I have been reading about nowcasting. From what I understand what UMIDAS does is that it transforms the dataset into cross sectional data and then runs OLS.

The more I read articles applying Random Forest, SVM and others in nowcasting the more I am understanding that cross sectional data is used. Am I wrong? If not is this not very bad. I have run some models to nowcast GDP cross sectioning Month 1,2 and 3 as my X's and GDP as my Y and it performs bad. Are they doing it different?

$\endgroup$
1
  • $\begingroup$ There is very little data to go on here, I have answered your question given what you have said. But more data would help, for example what does your data, and 'their' data, look like, which UMIDAS approach are you talking about? Who is doing it differently? What is there or your code? How could we assess these things without more info. $\endgroup$
    – Comte
    Apr 12 at 11:08

1 Answer 1

0
$\begingroup$

Nowcasting is simple modelling the recent past, present and very near future, which is what you will read on almost every article and blog about the subject. One does not need to use cross-sectional data, it's typically used because its available and relevant to application, such as filling in low frequency economic data with higher frequency data.

Nowcasting is also commonly used in meteorology, where forecasters get regular data updates from local sensors and satellites etc, this information, in time series, is received in real time. Its then used to make predictions, about the weather, either for that moment of time, or the very near future i.e. next 6 hours. This data could be made into a cross section, if we needed too, but the overall point here is that instead of simply focusing on updating forecasts, with new data coming in (streaming) we are focusing on say the next 6 hours. You do this to improve accuracy, and reduce error.

This short forecast could in turn, be used to fill in the data in longer streamed forecasts, if one found it to be useful (testing, testing, testing). And you could also turn the recent data (often high frequency) data, and/or predictions, into cross sectional data, that inform other forecasts. Would it help? Depends on the application and the data.

Regarding your specific use case, there is a lack of detail. Considering you are looking at an economic metric (GDP), I would say using cross-sectional data for nowcasting is appropriate. Without knowing the details of the data, I can only really suggest working on your feature selection. There are a lot of approaches to nowcasting gdp, which all depend variations in cross-sectional data, for example...

  1. atlantafed
  2. European central bank now casting with card data
  3. Fed global GPD nowcasting
  4. Bank of england

As you can see there isn't a single approach to modelling GDP, the fed paper discusses nowcasting in extreme situations, which is defiantly worth looking at, if you want to build a robust model. Its important to remember have a good backtesting approach (just covering the basics here). You have probably seen this UMIDAS approach, are they doing it wrong? Maybe, there is not github code or data to look at, model performance comes down to your data first.

hope that helps

note. I am just a forecasting data scientist not an economist.

$\endgroup$
2
  • $\begingroup$ So what I am asking is lets say I have GDP and a monthly variable. How do I build the dataset? Is it Jan's value for X,Feb's value for X and March's value for X and as Y the Q1 GDP then on the next row April's value for X, May's value for X, June's value for X and as Y Q2 for GDP. Or is it Nov,Dec,Jan against Q1 then Dec,Jan,Feb then jan, Feb,Mar on 3 different rows with Y's being Q1 GDP and on the next row Feb,Mar,April and Q2 GDP against it and so forth? If so, how to interpret prognosis? $\endgroup$
    – J_Bake
    Apr 13 at 12:33
  • $\begingroup$ Or if using sparse matrices what ML models can handle those? $\endgroup$
    – J_Bake
    Apr 13 at 15:19

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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