Tack forecasting

Model for forecast a tack event in sailing from sensor data. Interview task for T-DAB.

Requirements

• Python 3.8

The Python packages needed are listed in requirements.txt

Setup

Unix setup with virtualenv

virtualenv venv  
source venv/bin/activate  
pip install -U pip  
deactivate  
source venv/bin/activate  
pip install -r requirements.txt  

Data setup

mkdir data/
mkdir data/train/  
mkdir data/test/  
mv PATH/TO/test_data.csv data/  
cd src  
python split_train_test.py  

Model explanation

The tack forecasting model consists of a VAR model for forecasting the future state of the sailing boat and a random forest for classifying whether a tack is occurring in this state.

The assumptions made in designing this model were:

• Tacking is a function of the current state only.
• When not tacking, the state can be described by an auto-regressive process and therefore is wide-sense stationary.
• When tacking, the state can be made wide-sense stationary.

These assumptions were verified empirically before selecting and constructing a model.

Model evaluation

The model is composed of forecasting and classification components. The complete model can be evaluated using the eval_model.py script in the src directory. This script takes an optional argument with the path to a CSV file containing a single (labelled) sailing sequence to be used for evaluating the model accuracy. If no path is passed, a random sequence from the test dataset is chosen for evaluation. Usage:

python eval_model.py --datapath PATH/TO/DATA.csv

The classification model can similarly be evaluated (only on the test set, not custom data) by running:

python eval_classifier.py

API usage

The run_model.py script shows how to use the API to forecast from a new datapoint (note that a history of datapoints of a user-specified length is required for forecasting). The eval_model.py shows how this can be extended to an input stream of datapoints, as is the case in real-world application.