Welcome to my MSc Thesis for completion of the MSc Quantitative Finance at the Erasmus University Rotterdam. In this study, I simulate 1-day, 10-day, and 20-day short rate paths for EONIA using the TimeGAN model. Next to that, I evaluate the ECB's mapping of EONIA to €STER and check the applicability of TimeGAN for interest rate simulation of €STER, i.e. after the EONIA-€STER transition. See the Thesis or check out the rest below!
Below is the visualization of the results presented in Table 4.
T-VaR(99%) estimate for regular TimeGAN (left) and TimeGAN with PLS+FM (right) during the validation dataset.
Below is the visualization of the results presented in Table 4 and Table 5 for TimeGAN with PLS+FM.
1-day, 10-day, and 20-day VaR(99%) estimates for TimeGAN with PLS+FM during validation and test dataset.
- 4 Training TimeGAN
- For CPU version of TimeGAN, see scripts/tgan.py
- For Multi-GPU version of TimeGAN, see LISA/tgan.py
- Visualization of training with Tensorboard
- 5 Data
- To produce Figure 5 until 9, see plotting.py
- To produce Table 3, see stylized_facts.py
- 7.1 Model selection
- To produce Figure 11 until 14, see autoencoder_training.py and hyper_and_importance.py
- 7.2 Coverage test
- To produce Table 4, see TimeGAN_kupiec.py, kalman_filter_vasicek.py, and variance_covariance.py
- 7.3 Diversity of simulations
- To produce Figure 22 and 23, see TimeGAN_kupiec.py
- To produce Figures 24, 36 until 43 see main.py
- 7.4 ECB's proposed mapping
- To produce Figures 26 until 28, see metrics.py, stylized_facts.py and main.py
Would like to see how the model works? This web application shows the influence of different hyperparameters and allows you to generate your own EONIA or €STER simulations. Note that it is still under construction
To train the TimeGAN model on EONIA data, install the folder locally using npm and run tgan.py:
$ cd ../TimeGAN-short-rates
$ npm install tgan.py
$ python tgan.py
>> [step: 1, g_loss_u_e: 0.018, g_loss_s: 0.023, g_loss_s_embedder: 0.021, e_loss_t0: 0.312, d_loss: 0.014]
>> [step: 2, g_loss_u_e: 0.019, g_loss_s: 0.022, g_loss_s_embedder: 0.031, e_loss_t0: 0.314, d_loss: 0.029]
...
To simulate 20-day short rate paths of EONIA or €STER, run the following code in python
# Generic packages
import pandas as pd
from sklearn import preprocessing
import tensorflow
# TimeGAN specific functions
from metrics import load_models
from training import RandomGenerator
# Import data and apply min-max transformation
df = pd.read_csv("data/Master_EONIA.csv", sep=";")
df = df.iloc[:, 1:] # Remove the Date variable from the dataset
df.EONIA[1:] = np.diff(df.EONIA) # Make first difference for EONIA
df = df.iloc[1:, :] # Remove the first value
scaler = preprocessing.MinMaxScaler().fit(df) # Perform min-max transformation
# Load the models, simulate scaled short rates and unscale
load_epochs = 8250, hparams = [], hidden_dim = 4, T = 20, nr_simulations = N
_, recovery_model, _, generator_model, _ = load_models(load_epochs, hparams, hidden_dim)
Z_mb = RandomGenerator(N, [T, hidden_dim])
samples = recovery_model(generator_model(Z_mb)).numpy()
reshaped_data = samples.reshape((samples.shape[0]*samples.shape[1],
samples.shape[2]))
scaled_reshaped_data = scaler.inverse_transform(reshaped_data)
simulations = scaled_reshaped_data.reshape(((samples.shape[0],
samples.shape[1],
samples.shape[2]))) Project is created with:
- Tensorflow version: 2.2
- Python version: 3.6.0
- Tensorboard version: 2.2
- Plotly Dash 1.16
This MSc Thesis is inspired on TimeGAN by @jsyoon0823