Scaling Deep Learning Multivariate Forecasting with no&low code
At Predicto, we often need to train and experiment with dedicated deep learning models for different stocks and cryptocurrencies, using similar categories of features but different data for each symbol. To streamline this process, we developed the concept of Data Model Templates (DMTs). In this blog post, we’ll explain how DMTs can be used to quickly and easily train hundreds of models at scale, and how our low code API makes it all possible.
1/the no code step
To create a DMT, we use our no code platform to describe the model using a data source (database or csv). We select features, prediction column, lookback and lookahead horizon, and add placeholders for values that we’ll parameterize later via our no code API (e.g. symbol name). This process takes just a few minutes and results in a DMT like the one shown below (important parts highlighted):
More details on how to create a DMT can be found in this blog post Forecasting News Coverage
2/the low code step
Now we are ready to scale by applying the DMT we created to as many stocks/crypto we want. The only thing we have to do is to call our nocode api, fetch our DMT by its id, handle the replacements by providing the {SYMBOL}/{FROM_DATE} and finally send for training.
Done — In a few hours, we’ll have hundreds of deep learning models ready for forecasting thanks to our auto-scaling cloud docker training agents!
In a similar way, once training is complete, we can easily schedule and retrieve forecasts from all the models we trained on demand. The only thing we need to do is pass it a time_index key and the nocode platform will fetch necessary history data from database, load model and generate a forecast!
And that’s it!
Here are some forecasts that were generated using DMTs (screenshot taken directly from Predicto website):
Wrapping up
Our platform for training and deploying deep learning models for stock and cryptocurrency forecasting has proven to be a powerful and efficient tool. By using data model templates and a low code API, we’re able to quickly and easily train and deploy hundreds of models, providing valuable forecasts for our users.
If you have any questions, feel free to get in touch or comment!
