2

I usually teach finance (asset pricing and equilibrium models), quantitative economics (linear algebra and optimization), econometrics, computer science introduction to programming and machine learning (Bishop and Bengio). I am preparing a 60-hour introductory course of time series for PHD students. I would like to mix theoretical and computational issues. Although the student can use any computer language he wants, I will provide all the examples in python.

There are some expected topics that cannot disappear from any time series introductory course for a department of economics:

  1. ARMA models

  2. Prediction and impulse response functions based on ARMA models

  3. Representations (covariance, Wold and spectral)

  4. Unit roots

  5. Cointegration

However, there are others that I am not exactly sure what to include:

  1. Forecasting (machine learning approach?)

  2. Multivariate models (VAR?)

  3. Nonlinear models (??) [TAR (?), Markov Switching (?), Garch Models (?)]

  4. Topics of Monte Carlo and Bootstrapping for time series (?)

  5. Contemporary issues? State of Art issues?

  6. Correct balance between theory and computation.

Can you help me suggesting topics (and references for these topics) that I could include in the course? Are there time-series references such as Wooldridge (Introductory Econometrics: A Modern Approach) or Inzenman (Modern Multivariate Statistical Techniques: Regression, Classification, and Manifold Learning) that use real data that can be used to exemplify the models?

kjetil b halvorsen
  • 63,378
  • 26
  • 142
  • 467
DanielTheRocketMan
  • 1,400
  • 11
  • 20
  • 5
    It would help if you were to explain what you mean by 'great data'. If I were to add a topic to your list it would be to discuss panel data models. There are many introductions to this topic but Pesaran's *Time Series and Panel Data Econometrics* is one of the best and most comprehensive texts. –  Sep 07 '19 at 14:20
  • 1
    Edited. Please: great data (real data that can be used to exemplify the models) – DanielTheRocketMan Sep 07 '19 at 14:36
  • 2
    Adding to @user332577 I would add Banerjee, A., Dolado, J. J., Galbraith, J. W., & Hendry, D. F. (1993). *Co-integration, error correction, and the econometric analysis of non-stationary data*. Oxford University Press, USA. – Alexis Sep 07 '19 at 15:43
  • 1
    @DanielTheRocketMan I think a reasonable way to approach this is to look at what other people are teaching at PhD level for economics. For example, see [here](https://www.eui.eu/DepartmentsAndCentres/Economics/Courses/Compulsory-Courses-2019-2020) the European University Institute econometrics course outlines. Including VARs is a no brainer and for macro you'll need to cover rational expectations models and the state-space form. I think you need to target the material towards economics too... if it's just time-series from a statistics angle, that's not much use for economics. – Graeme Walsh Sep 08 '19 at 00:47
  • 1
    @DanielTheRocketMan On "use real data that can be used to exemplify the models?" - I think a good approach here would be to use actual papers / models from economics journals as the examples and have students try to replicate them (after you've tried). Again, here the focus should be on the economics... and not losing focus of that in favour of programming or statistics. – Graeme Walsh Sep 08 '19 at 01:01
  • @GraemeWalsh "Including VARs is a no brainer" you mean that should not be included? I took a look in the link that you suggested and it is a topic 4 of https://www.eui.eu/Documents/DepartmentsCentres/Economics/Researchandteaching/Courses/CourseOutline-TimesSeriesAndSimulation-Based-Estimators-Cooper-Bueren2019.pdf – DanielTheRocketMan Sep 08 '19 at 01:18
  • 1
    @DanielTheRocketMan I meant that VARs should definitely be included. – Graeme Walsh Sep 08 '19 at 09:38
  • 1
    @user332577 Caveat: I have been reading Pesaran's text, and am finding errors in formulas (well, Ok, it is a 1st edition), and also questionable omissions (e.g., the formula provided for Spearman's rank correlation coefficient is *only* applicable when there are no ties, but Pesaran does not mention this). – Alexis Sep 24 '19 at 16:54
  • 1
    @alexis Fair enough. You should share these caveats with Pesaran himself as they are grist for a potential next edition. That said, before Pesaran, when deploying a panel data model one had to assume away issues such as weak dependence between cross sections, stationarity, autocorrelation, and so on. Pesaran's contributions to PDMs wrt integration of classic time series issues, in my view, far outweighs any peccadilloes in his text. –  Sep 26 '19 at 14:04

1 Answers1

1

Non sensical results from auto_arima is one example of real data presenting real challenges to model-presumptive approaches. I have contributed a ton of responses including data with results/discussions that might be useful to you as you bridge the gap between theory and practice ....which is (I am sure) what you are trying to do.

IrishStat
  • 27,906
  • 5
  • 29
  • 55
  • Yes, it is exactly what I am looking for. I will take a look in your answers! – DanielTheRocketMan Sep 07 '19 at 19:15
  • When I used to teach statistics (I have not done this for a while) I loved to present some curious real world problems that could specially call the attention of the students such as the Simpson Paradox, Monty Hall Paradox, Benford Law, Secretary Problem, The Coupon Collector's Problem and coincidences (the birthday problem and the north caroline lottery coincidence)... Curious cases in time series would be very welcome. – DanielTheRocketMan Sep 07 '19 at 19:27
  • 1
    why don't you contact me offline and I wil try to continue a dialogue .. I hate monologues. – IrishStat Sep 07 '19 at 19:32
  • I am not sure how to that (?). As far as I know I am not able to send a direct message to another user (?). – DanielTheRocketMan Sep 08 '19 at 00:39
  • use the email address in my network profile . everybody has a profile . – IrishStat Sep 08 '19 at 01:30
  • 1
    From an economics standpoint, auto_arima functions, and arima models more generally, are of little interest. These models lack economic theory, so they’re not useful to an economist. Moreover, they don’t allow for model discovery, from an LSE gets perspective. Arima models don’t allow the economist to distinguish him/herself from a statistician. For that reason, I wouldn’t spend much time on them in a course delivered by a school of economics. You need to ask what sort of economists are you trying to train and tie the course in with whatever is being taught in the micro and macro courses. IMO. – Graeme Walsh Sep 08 '19 at 15:00