Most of economic data are time series. Regresssion of time series can lead to spurious regression.What are the main issues while analyzing time series?In other words, what are the main statistical tests to be performed before analyzing the time series, such stationarity test?
https://thephysicspi.blogspot.com/2019/02/advantages-and-disadvantages-of-time.html
https://www.researchgate.net/post/What_are_the_weaknesses_of_time_series_model_and_how_could_it_be_improved
https://www.dummies.com/education/economics/econometrics/how-to-deseasonalize-time-series-data/
-------------------------------------------------------------
The understanding and 'use' of time (dominantly as linearity) itself in economics is the key issue, in methodological terms. With respect to this point, the 3rd link is essential, in my humble opinion.
---------------------------------------------
https://plato.stanford.edu/entries/time-thermo/
https://en.wikipedia.org/wiki/Entropy_(arrow_of_time)
https://thephysicspi.blogspot.com/2019/02/advantages-and-disadvantages-of-time.html
https://www.researchgate.net/post/What_are_the_weaknesses_of_time_series_model_and_how_could_it_be_improved
https://www.dummies.com/education/economics/econometrics/how-to-deseasonalize-time-series-data/
-------------------------------------------------------------
The understanding and 'use' of time (dominantly as linearity) itself in economics is the key issue, in methodological terms. With respect to this point, the 3rd link is essential, in my humble opinion.
---------------------------------------------
https://plato.stanford.edu/entries/time-thermo/
https://en.wikipedia.org/wiki/Entropy_(arrow_of_time)
It depends on the aims of your research. What relates to your research:
A) to produce a reasonably accurate forecast
B) to study the features of your series (such as the presence of trend or seasonality)
C) to test hypotheses that one variable is affected by another variable
Case A: Yes, stationarity tests to select a good model. Then apply residuals tests (to detect the presence of heteroscedasticity and problems with error term distribution), to see if your forecasting model is adequate.
Case B: Tests for trend (such as Mann-Kendall) and for seasonality can be applied.
Case C: Stationarity, causality test + if you want to have a parametric model, all the tests from case A.
P.S. The above link sent by Stephen I. Ternyik is great. One important issue, of course, is that you can obtain a very unreliable forecast if you make a too strict assumption that the dependencies observed in past will hold true in future. On the other hand, if your model is too flexible, you can obtain a forecast with a very high level of uncertainty. One solution here is to try to integrate judgment and statistical methods. Some solutions can be found in my thesis:
Thesis Integration of judgmental and statistical approaches for dem...
If you want to compare alternative forecasting models across many series, you can use the statistical test described in this chapter (page 23):
Chapter Measuring Forecasting Accuracy: Problems and Recommendations...
Hi,
I am in the same situation right now. As a psycholgist, I do not know very much about time series. What had helped me (I am doing this in R) is
1) the tidyverse approach by the "tidyverts" (Rob Hyndman, Mitchel O'Hara-Wild and Earo Wang) and especially the R packages "tsibble", "fable", and "feasts").
This extremely facilitates learning time series analysis. By searching those names on youtube you will find some nice videos. Also, there are tutorials that you find by searching.
2) There is an online book by Rob Hyndman:
Hyndman, R. J., & Athanasopoulos, G. (2018). Forecasting: principles and practice: OTexts.
https://otexts.com/fpp2/
3) most of the lit on time series seems to be focused on the forecasting and descriptive goal of time series analysis (as Andrey has pointed out). This was/is very confusing for me as a researcher with explanatory goals. In this regard I started this week reading the book by Box-Steffensmeier et al. which is very much focused on estimating models with predictors. This is an eyeopener for me.
Box-Steffensmeier, J. M., Freeman, J. R., Hitt, M. P., & Pevehouse, J. C. (2014). Time series analysis for the social sciences: Cambridge University Press.
Hope this helps
Holger
If the data fits your assumptions models and if there is variable correlation, then a time series is appropriate to use.
Monica, the problem I have is that I have time series data from 1991 to 2018, when I use OLS, all assumptions are respected. However, I m afraid of having spurious regression. When I test stationarity, all data have unit root. But, they do not have same lag. Some of them have l(1), others l(2) or l (3).
Sule thank you for contribution.Not considering structural breaks leads to a model which is not accurate.
IT DOES NOT WORK. Although it sounds impressive and all the equation looks so technical, in reality it is the most ineffective tool. it is is useful, it would be able to predict financial crisis and economic depression.
WHAT IS WRONG WITH TIME SERIES? It relies heavily on the past and regressing against itself, hence the term "autoregressive". The market and the economic are multifaceted, i.e. influence by many factors. Time series do not take that into account. The verbiage "let everything else remained constant..."---does not work in reality. "Time series method needs a reality check.
ECONOMETRIC does not go far because we are stuck with time series. Explore other methods. See here for a start: https://en.wikipedia.org/wiki/Mathematical_model
Dear Paul, thank for you you comment. I share with you your idea. Economic phenomenon are connected, one time series can not predit the future of it self. For instance , demand is in function with the revenue of the consumer, inflation has impact on the revenue,etc.
Times series have much limitations in termes of predicting by auto regressive.
Exactly: if we extrapolate past patterns, the forecast becomes inadequate. If we use very flexible models, the forecast becomes very diffuse. One way to overcome this problem is that we collect judgmental estimates first and then update forecasts as new data becomes available. Let me recommend this paper that addresses the issue of insufficient data and proposes a method to integrate judgment in a convenient way:
Technical Report A joint Bayesian forecasting model of judgment and observed data
I agree with Dr. Davydenko in that there is a way to deal with your problem. You may collect judgmental estimates first and then update forecasts as new data becomes available.
So if there is different order of integration like I(0), I (1) or I(2) , you may use ARDL not OLS
- Badges
- Science topic
- Similar topics
- Statistics
- Time Series