Numerauto

Numerauto is a Python daemon that facilates automatic weekly participation in the Numerai machine learning competition (http://numer.ai).

Users can implement custom event handlers to handle new training data, apply models to new tournament data when a new round starts, and more. Example event handlers are included for training SKLearn classifiers and uploading predictions.

If you encounter a problem or have suggestions, feel free to open an issue.

Installation

To install the latest release:

pip install --upgrade numerauto

If you prefer to use the latest development version, clone this github reposistory:

git clone https://github.com/thebrain85/numerauto

And then run the following command in the numerauto directory.

python setup.py install

Usage

Example

Although Numerauto can be run without any event handlers for testing, it will not do anything except detecting round changes and downloading the new dataset every week.

See example.py for a basic example that trains a scikit-learn logistic regression model and uploads its predictions.

The example uses the PredictionUploader event handler to upload predictions to Numerai. This requires you to register an API key in your Numerai account. This can be done in Account settings -> Your API Keys -> Add. Select 'Upload submissions' to be able to upload predictions using this API key. Replace 'insert your publickey/secretkey here' in the code with your own API public/secret API key pair to allow the example code to upload the prediction to your account.

See example2.py for an example that uses the CommandlineExecutor event handler to call a custom commandline once a new round is detected. You can modify the command line to execute your own code, e.g. python myscript.py, like you would do manually. This is an easy and quick way to make use of Numerauto's automatic detection of new rounds in Numerai.

Custom event handlers

Implementing your own event handler is easy. Simply create a subclass of numerauto.eventhandlers.EventHandler and overload the on_* methods that you need to implement your own functionality. The event handler can then be added to a Numerauto instance using the add_event_handler method. Then start the main loop of the Numerauto daemon by calling the run method.

Currently these events are supported:

• def on_start(self): Called when the daemon starts.
• def on_shutdown(self): Called when the daemon shuts down.
• def on_round_begin(self, round_number): Called when a new round has started.
• def on_new_training_data(self, round_number): Called when the daemon has detected that new training data is available.
• def on_new_tournament_data(self, round_number): Called every round to signal that there is new tournament data.

Note that event handlers are called in the order they are added to the Numerauto instance. Also note that all handlers for one event are called before the next event is handled, keep this in mind when designing event handlers that interact with one another, or that keep large amounts of data in memory. Ideally, the handler should clean up memory in on_new_tournament_data to prevent memory being used while the daemon is idle and waiting for the next round.

Running Numerauto

By default, the run method of Numerauto will keep running indefinitely until interrupted using a SIGINT (ctrl-c) or SIGTERM signal. This way, you only have to start your Numerauto script once, for example on startup, or running inside screen on linux. example.py runs Numerauto this way.

Alternatively, you can provide the argument single_run to the run method of Numerauto. This will cause Numerauto to only wait once for a new round (and with a maximum of 24 hours). This is ideally suited for calling Numerauto from a task scheduler, such as cron on Linux or the Windows Task Scheduler. Make sure you schedule Numerauto to run a little before the official round start time, it will wait and run as soon as the dataset is available. example2.py runs Numerauto this way.

Persistent state: state.pickle

Numerauto stores a persistent state in the state.pickle file in the directory from which the daemon is being run. By default, the Numerauto daemon stores the last round number that was processed (last_round_processed) and the last round number on which training was performed (last_round_trained). You can force the system to reprocess and retrain by stopping the daemon and removing the state.pickle file.

Custom event handlers can store persistent information in the persistent_state dictionary of the Numerauto instance.

Round report

Event handlers have access to a special dictionary in the numerauto instance via self.numerauto.report. This dictionary is reset every round and can be used to report values that are relevant to the event handler. For example, SKLearnModelTrainer reports back the filenames of the trained model (if applicable) and the generated predictions, and PredictionStatisticsGenerator reports the validation metrics of a prediction.

The report can be written to file every round with BasicReportWriter, or emailed with BasicReportEmailer, both using only simple formatting.