Team Aleph Omega Analytics solution (John Lyons, Florian Weiß, Max Weber, Benedict Schwind)

Trading Hub Europe Challange

This repository contains:

• Training data used for competition
• Source code to preprocess data and make prediction
• Source code to impute some of the features which can lead to an improvement of the score (rlm_imputation, exit_imputation, entry_imputation)

How to run

• Install python requirements from requirements.txt
• Open the corresponding notebook for either 1-day or 4-day-ahead prediction
• Tweak parameters in "model prediction area". "repeats" controls how often to look for best epochs and best seed for current slot prediction (using a mini val which is 25 days before prediction slot). You can also control whether or not do training for every slot (i.e. only train for every nth slot)
• Run each cell sequentially

Code notes

The code runs smoothly, with the exception that it is difficult to find the right code locations for appropriate configurations. This is due to the fact that there are exams at the TH Rosenheim at the deadline and therefore only little time was left for cleaning up the code.

The gridsearch script is not at the same level as the prediction generation scripts. This is because it has not been used since it figured out the best model for the task at hand. However, it may be used in future work to try out different combinations.

How to generate rlm_imputation, exit_imputation, entry_imputation

Our main forecast model uses some extra features like rlm_imputation, exit_imputation, entry_imputation. Imputation means in this context, that this features are forecasted by another model to boost the performance of the main model. These extra features are already merged into train.csv which will be used if you execute the submission notebooks of the main model (neural_network). For reproducibility and transparency, there are three notebooks which shows how we generate these extra features.

You simply need to execute these notebooks to get several csv files which contains the imputed/forecasted values. Please check, if the code iterate over all slots and the csv are exported properly. Further, for a full examaple of what we did, it will be neccessary to merge these new imputed features into data/train.csv (and overwrite the old values which were already meged by us).

DB Regio Bus Challenge

This repository contains:

• Training data used for competition
• Source code to preprocess data and make prediction
• Scientific Paper

How to run

Unzip the data folder an run the two R scripts to generate two csv files which contains the forecasts for on-demand/regular task.

To reproduce the data preparation steps, unzip raw data in 30data/db_regio_bus/data.zip and execute the jupyter notebooks and provide