Dev Update – 2nd Dec 2020

Last week, we had Corona infections in my office building, and we got a shutdown for two days. As winter has come, The corona situation in Korea is getting serious. The number of infections increased, and many places are being closed. All team members are safe so far.

James delivered a presentation at Webworld 2020 conference(http://www.bizdeli.com/webworld/), talking about Future trend of Defi. And he gave a lecture, how machine learning can add value to the financial industry to the Korean Intellectual Property Office.

As for Orbit, We have launched Orbit v2 this week. After monitoring the activity of Orbit v2 and Gaia(Automated asset management system), we will put more money into it. We hope that we do not see any severe bugs in our system. We all are excited about the Orbit v2. 🙂

 

Athena

Orbit v2

  • We added a sibling model of Orbit v2 for testing. Thanks to recent update, Orbit v2 has customizable parameters such as data preprocessing, feature engineering, and so on.
  • Missing data handling routine is updated. Orbit v2 model processes data on the fly, retrieving data from database and do some data processing. Due to some reason, We do not use preprocessed data. As the number of the model and instruments increases, the probability of missing data and wrong data results is high. This is a critical error because the machine learning model expects the exact same data input for prediction. If the data input is different, it does not generate an outcome. It took some time to update the code.

Mercury

Database Tuning

  • As the number of models and instruments is bigger than a test environment, database performance is degraded. Orbit and other components in Gaia request a large amount of data to database more than few gigabytes at once. This puts a significant burden on the database system. So database access time could reach more than 10 minutes in the worst case.
  • To resolve this performance issue, we optimized database parameters, increasing memory buffer, throughput related parameters, and so on. We know that this is not a permanent solution. But it is good for now.

Deployment to Production Environment

  • We have three environments for Gaia, Development, Test, and Production. All these environments are identically the same except its purpose. Since Gaia has many components, deployment is a complicated process. Checking deployment is another process with extreme care.
  • All components have been deployed in kubernetes environment. We exposed only mandatory components to the internet for security reasons. Now Gaia is in the production environment.

Get early access to the latest updates on the MoA Investment Solutions