EconDL v1.1 (15 Sep 2022)

Instructions:

• Change your experiment parameters by creating a new json file like {run_name}.json in the exp_config folder. Make sure not to leave out any parameters (look at the example, 18jul_test.json)

  • For nn_hyps that don't change throughout experiments, they are defined in the nn_hyps_default.json file or wherever you set the default_nn_hyps parameter to be. You can also throw those parametes into run_params (e.g. for 18jul_test.json, all experiments have the same constant_tvpl, so I put them in run_params)

  • NEW 9/15 Now instead of specifying s_pos, we only need to specify s_pos_setting, which will automatically calculate the dimensions for you under the hood. The possible settings are:

    • endog - only the VARNN endogenous variable lags will be added
    • endog_time - 2 hemispheres: 1) VARNN endogenous lags 2) Time (trends or dummies depending on time_dummy_setting)
    • endog_exog - 2 hemispheres: 1) VARNN endogenous lags 2) Exogenous variables of the dataset
    • combined - endogenous, exogenous and time variables in ONE hemisphere (not really recommended as different number of variables imposes an incorrect prior on importance of each type of variable)
    • exog - only the exogenous data (not really used)
    • time - only the time hemisphere

  • Specify s_pos_setting like this: "s_pos_setting": {"hemis": "endog"}

  • Remember that each experiment's nn_hyps dictionary will overwrite everything in run_params (which are supposed to be parameters that apply to the entire set of experiments). Hence, if you want to run endog, endog_time etc. in different expeirments, then specify s_pos_setting in each experiment's dictionary in nn_hyps.

• To run the experiment, run python execution_script.py {run_name} {num_experiments} in command-line. (On Compute Canada, put this command in the .sh file and run sbatch 18jul_test.sh as before)

  • It is important that run_experiments is correct i.e. the number of elements in the nn_hyps error in your experiment json file. This is so that the code can correctly divide the Compute Canada cores to calculate different repeats and experiments properly.
    • e.g. If there are 4 experiments and I want to run 3 repeats per experiment, I will specify 1-12 in the Compute Canada .sh file. This means that cores 1, 5, 9 will run experiment 0, cores 2,6,10 will run experiment 1, ...

• Evaluation: When all the cores are done, run python evaluation_script.py {run_name} in command-line (or via .sh file in Compute Canada) to run the Evaluation object which plots all the graphs.

  • If some repeats failed and you want to exclude them from the evaluation, you can use the repeats_to_include parameter in the json file to specify the repeats you want to include.

Parameters to Note

• constant_tvpl - number of neurons in the TVPL layer for the constant. Must be wrapped by a list i.e. [40] (added 7/18)

Sequence of Events

Training (execution_script.py)

• For each run:
  • train_experiments() - For each experiment
    • For each repeat:
      • train_experiment()
      • compute_unconditional_irf()
      • compute_multi_forecasting()
  • train_multi_forecasting_benchmarks()
  • train_benchmarks()
  • For each ML experiment:
    • For each repeat:
      • train_experiment()
      • compute_unconditional_irf()
      • compute_multi_forecasting()

Evaluation (evaluation_script.py)

• For each experiment:
  • load_results - loads uncompiled results into self.results_uncompiled (only repeats specified in ‘repeats_to_include)
  • compile_results:
    • For VARNN Experiments: results, unconditional IRF - and plots it, multi forecasting
    • For ML Experiments: unconditional IRF - and plots it, multi forecasting
  • compute_conditional_irf (needs compiled results) - computes and plots conditional IRF results
  • plot_all() - plots VAR IRFs, cholesky, precision, sigmas, betas, one-step predictions, OOB and test errors, multi-step forecasts

Version History

• v1.1 (7/20/22): Self-tuning ML Benchmarks, Running experiments w multiple hemispheres
• v1.0 (7/19/22): Experiment and Evaluation code works - verified with 18 Jul example.
• What doesn't work/not implemented: prior shift, simulations evaluation, VSN

File Structure

• Run
  • Experiment
    • Repeat
• Classes:
  • Run (run-level)
  • Experiment (experiment-level)
  • Evaluation (run-level)
  • Benchmarks (run-level)
  • ForecastBenchmarks (run-level)
  • IRFConditional (experiment-level)
  • IRFUnconditional (repeat-level)
  • IRFUnconditionalEvaluation (run-level)
  • ForecastMulti (repeat-level)