TimeSeriesGAN

Time series analysis using GANs

The model in this repository combines a stack of dilated causal convolutional and LSTM layers, and while GANs are quite challenging to train, this approach have produced great results on various types of datasets, such as stock market and weather measurements.

├── data
│   └── sample.csv
├── databricks-ci
│   ├── create_cls.sh
│   ├── runNotebook.sh
│   └── start_cls.sh
├── .gitlab-ci.yml
├── Standalone_example.ipynb
└── TSGAN_Hypertuning.ipynb

Requirements

• Versions are just recommendation

Databricks ML runtime (e.g. 10.5.x-cpu-ml)
Optuna 3.0.1
MLflow 1.8.0
torch 1.5.0
torchvision 0.6.0

Data extraction:

• The following script uses Spark for fast extraction of the latest financial data from Yahoo Finance API

tickers = ["GOOG", "AAPL"]
df = spark.createDataFrame([(i,) for i in tickers],("ticker",))

return_type = ArrayType(MapType(StringType(), StringType()))

@udf(returnType=return_type)
def yahoo_udf(ticker_label):

    headers={'User-Agent': 'Mozilla/5.0 (Macintosh; Intel Mac OS X 10.12; rv:74.0) Gecko/20100101 Firefox/74.0'}
    end_ts = int(time.time())
    start_ts = end_ts - 3600*24*7
    ticker = ticker_label
    data_url = f"https://query1.finance.yahoo.com/v8/finance/chart/{ticker}?symbol={ticker}&period1={start_ts}&period2={end_ts}&useYfid=true&interval=1m&includePrePost=true"
    values = None
    data_dict = {i: [] for i in ['timestamp', 'low', 'high', 'close', 'open', 'volume']}
    try: 
        f = urllib.request.urlopen(urllib.request.Request(data_url, headers=headers))
        status = f.status
        if status == 200:
            jresp = json.loads(f.read().decode("utf-8"))    
            data_dict['timestamp'] = jresp['chart']['result'][0]['timestamp']
            for k in ['low', 'high', 'close', 'open', 'volume']:
                data_dict[k] = jresp['chart']['result'][0]['indicators']['quote'][0][k]
    except:
        pass
    df = pd.DataFrame(data_dict)
    values = df.to_dict("index").values()
    return list(values)

extracted = yahoo_udf("ticker")
exploded = explode(extracted).alias("exploded")
expanded = [
    col("exploded").getItem(k).alias(k) for k in ['timestamp', 'low', 'high', 'close', 'open', 'volume']
]

pdf = df.select("ticker", exploded).select("ticker", *expanded)

Hyperparameter optimization

• Optuna and Mlflow were through case studies and expirements used to provide the optimal choices for learning rate, optimization algorithm, dropout ratio for the Generator and the Discriminator independently

def suggest_hyperparameters(trial):
    lr = trial.suggest_float("lr", 1e-4, 1e-1, log=True)
    dropoutG = trial.suggest_float("dropoutG", 0.0, 0.4, step=0.1)
    dropoutD = trial.suggest_float("dropoutD", 0.0, 0.4, step=0.1)
    optimizer_name = trial.suggest_categorical("optimizer_name", ["Adam", "Adadelta"])

    return lr, dropoutG, dropoutD, optimizer_name

def objective(trial):
    best_val_loss = float('Inf')
    nz_dim = 100
    best_mse_val = None
    ...
    with mlflow.start_run():

        lr, dropoutG, dropoutD, optimizer_name = suggest_hyperparameters(trial)
        ...
        mlflow.log_params(trial.params)
        mlflow.log_param("device", device)

        netG = Generator(in_dim=nz_dim, n_channel=10, kernel_size=8, out_dim=1, hidden_dim=100, dropout=dropoutG).to(device)
        netD = Discriminator(in_dim=1, cnn_layers=4, n_layers=1, kernel_size=8, n_channel=10, dropout=dropoutD, hidden_dim=100).to(device)

        optimizerD = getattr(optim, optimizer_name)(netD.parameters(), lr=lr)
        optimizerG = getattr(optim, optimizer_name)(netG.parameters(), lr=lr)
        
        train(netG, netD, device, train_dataloader, optimizerG, optimizerD, n_epochs)
        mse_errG = test(netG, device, test_dataloader)
        
        if best_mse_val is None:
            best_mse_val = mse_errG
        best_mse_val = min(best_mse_val, mse_errG)
        mlflow.log_metric("mse_errG", mse_errG)

    return best_mse_val


run_tag = datetime.datetime.now().strftime("%Y-%m-%dT%H:%M")
experiment_id = mlflow.create_experiment(
    f"/Users/user/TimeSeriesGAN_Exp_{run_tag}",
    tags={"version": "v1", "priority": "P1"},
)

mlflow.set_experiment(experiment_id=experiment_id)
study = optuna.create_study(study_name=f"TimeSeriesGAN_study_{run_tag}", direction="minimize")
study.optimize(objective, n_trials=10)

Results

sample.csv

• D(G(z))

Evolution of the generated time series

• Epoch 67

• Epoch 135

• Epoch 200