Skip to main content

A recurrent neural network for predicting stock market performance

Project description

AlphaNet

unittest Congyuwang publish

A Recurrent Neural Network For Predicting Stock Prices

AlphaNetV2

Below is the structure of AlphaNetV2

input: (batch_size, history time steps, features)

              stride = 5
input -> expand features -> BN -> LSTM -> BN -> Dense(linear)

AlphaNetV3

Below is the structure of AlphaNetV3

input: (batch_size, history time steps, features)

                 stride = 5
        +-> expand features -> BN -> GRU -> BN -+
input --|       stride = 10                     |- concat -> Dense(linear)
        +-> expand features -> BN -> GRU -> BN -+

Installation

Either clone this repository or just use pypi: pip install alphanet.

The pypi project is here: alphanet.

Example

Step 0: import alphanet

from alphanet import AlphaNetV3, load_model
from alphanet.data import TrainValData, TimeSeriesData
from alphanet.metrics import UpDownAccuracy

Step 1: build data

# read data
df = pd.read_csv("some_data.csv")

# compute label (future return)
df_future_return = here_you_compute_it_by_your_self
df = df_future_return.merge(df,
                            how="inner",
                            left_on=["date", "security_code"],
                            right_on=["date", "security_code"])

# create an empty list
stock_data_list = []

# put each stock into the list using TimeSeriesData() class
security_codes = df["security_code"].unique()
for code in security_codes:
    table_part = df.loc[df["security_code"] == code, :]
    stock_data_list.append(TimeSeriesData(dates=table_part["date"].values,                   # date column
                                          data=table_part.iloc[:, 3:].values,                # data columns
                                          labels=table_part["future_10_cum_return"].values)) # label column

# put stock list into TrainValData() class, specify dataset lengths
train_val_data = TrainValData(time_series_list=stock_data_list,
                              train_length=1200,   # 1200 trading days for training
                              validate_length=150, # 150 trading days for validation
                              history_length=30,   # each input contains 30 days of history
                              sample_step=2,       # jump to days forward for each sampling
                              train_val_gap=10     # leave a 10-day gap between training and validation

Step 2: get datasets from desired period

# get one training period that start from 20110131
train, val, dates_info = train_val_data.get(20110131, order="by_date")
print(dates_info)

Step 3: compile the model and start training

# get an AlphaNetV3 instance
model = AlphaNetV3(l2=0.001, dropout=0.0)

# you may use UpDownAccuracy() here to evaluate performance
model.compile(metrics=[tf.keras.metrics.RootMeanSquaredError(),
                       UpDownAccuracy()]

# train
model.fit(train.batch(500).cache(),
          validation_data=val.batch(500).cache(),
          epochs=100)

Step 4: save and load

saving

# save model by save method
model.save("path_to_your_model")

# or just save weights
model.save_weights("path_to_your_weights")

loading

# load entire model using load_model() from alphanet module
model = load_model("path_to_your_model")

# only load weights by first creating a model instance
model = AlphaNetV3(l2=0.001, dropout=0.0)
model.load_weights("path_to_your_weights")

Note: only alphanet.load_model(filename) recognizes custom UpDownAccuracy. If you do not use UpDownAccuracy, you can also use tf.keras.models.load_model(filename).

Documentation

For detailed documentation, go to alphanet documentation.

For implementation details, go to alphanet source folder.

One Little Caveat

The model expands features quadratically. So, if you have 5 features, it will be expanded to more than 50 features (for AlphaNetV3), and if you have 10 features, it will be expanded to more than 200 features. Therefore, do not put too many features inside.

One More Note

alphanet.datamodule is completely independent from alphanet module, and can be a useful tool for training any timeseries neural network.

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

alphanet-0.0.20.tar.gz (17.2 kB view hashes)

Uploaded source

Built Distribution

alphanet-0.0.20-py3-none-any.whl (16.3 kB view hashes)

Uploaded py3

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page