wcswcswcs / fts-gan

FTS-GAN

This package is an implementation of the FTS-GAN framework from "FTS-GAN: Generating Multidimensional Financial Time Series". The framework allows the generation of synthetic data by learning an underlying distribution.

Getting Started

Data can be loaded from a csv using one of the data loaders. If the S&P 500 and VIX is used as an example to load a csv the following is done:

from ftsgan.ftsgan import load_data

dataset = load_data(
        file_dir=data_dir + 'sp500_vix.csv',
        columns=['S&P 500', 'VIX'],
        window_size=255,
        start_date='1992-01-02',
        end_date='2020-12-31',
        format='%Y/%m/%d',
        device=device,

To setup the model the parameters must be passed as well as the metrics being used to monitor the quality of training.

from ftsgan.ftsgan import FTSGAN
from ftsgan.evaluation import ACFLoss, HistoLoss, KurtosisLoss, LevEffLoss, SkewnessLoss, CorrelationLoss, SpreadLoss
params = {
    'g_in_channel': 6,
    'g_hidden_layers': [80] * 7,
    'g_skip_layer': 80,
    'd_hidden_layers': [80] * 7,
    'd_skip_layer': 80,
    'objective_features': ['acf_abs', 'acf_id', 'kurtosis'],
    'topk': 200
}
sp_data = dataset.series['S&P 500'].data.permute(0, 2, 1).contiguous()
sp500_metrics = [
    HistoLoss(sp_data, n_bins=200, name='abs_metric'),
    KurtosisLoss(sp_data, name='kurtosis'),
    SkewnessLoss(sp_data, name='skewness'),
    ACFLoss(sp_data, max_lag=64, name='acf_id', threshold=0.3),
    ACFLoss(sp_data, max_lag=64, name='acf_abs', transform=torch.abs, threshold=0.3),
    LevEffLoss(sp_data, name='lev_eff'),
]
vix_data = dataset.series['VIX'].data.permute(0, 2, 1).contiguous()
vix_metrics = [
    HistoLoss(vix_data, n_bins=200, name='abs_metric'),
    KurtosisLoss(vix_data, name='kurtosis'),
    SkewnessLoss(vix_data, name='skewness'),
    ACFLoss(vix_data, max_lag=64, name='acf_id', threshold=0.3),
    ACFLoss(vix_data, max_lag=64, name='acf_abs', transform=torch.abs, threshold=0.3),
    LevEffLoss(vix_data, name='lev_eff'),
]
all_metrics = [
    CorrelationLoss(sp_data, vix_data, name='corr_loss', threshold=0.04)
]
test_metrics = {0: sp500_metrics, 1: vix_metrics, 'all': all_metrics}
num_series = 2
batch_size=256
optimal_params = 'save'

model = FTSGAN(test_metrics, num_series=num_series, num_epochs=3000, batch_size=batch_size, params=params, save_dir=save_dir, ngpu=ngpu, verbose=5)

The optimal_params sets if the model will stop when a threshold is reached or will run for the set number of epochs and then save the best performing model.

The params overwrite the default implementation. The metrics must be chosen and passed to the test_metrics dictionary where each series is assigned a number which is used as the key.

To train the model it is as simple as calling the train method on the FTSGAN class:

model.train(financial_series, optimal_params=optimal_params)

Once training is complete samples can be generated by the following:

t_dim = dataset.series['S&P 500'].raw_data.shape[-1] - model.net_g.receptive_field_size()
synthetic_returns, synthetic_data, synthetic_paths  = model.generate_inverse(
    100,
    t_dim,
    dataset,
    optimal_params='save',
    pre_transform_dims={'S&P 500': (0, 2), 'VIX': (2)},
    accept_reject=False)

This will generate 100 samples of length given by t_dim.