Ethereum Transformer Bot

• Can Natural Processing Transformers be used for basic signal processing?

INTRO

Second experiment testing if natural language model can predict arbitrary binary signal movement. In our previous experiment, we needed to host our own GPT-2 model. Already a pretty massive file and growing, this model was too large to host on github and made for a confusing setup. It also required high minimum compute and memory to host and run any language task.

With GPT-3 comes a new paradigm of pay per call cloud transformer models. Using https://app.inferkit.com we set up an experiment which interprets Ethereum price movement and continuously feed it to the API. The bot's ability to predict the market's next move is logged to the database below.

Flow

Ethereum Price

• Market movement is converted to binary's based on increase or decrease

• Use the transformer API to extract opinion

• Average of 10 samples

Results

Average 53.8%

100hr blocks

0 - 100: 53%

200 - 300: 54%

300 - 400: 63%

400 - 500: 51%

500 - 600: 47%

• Python
• Talk To Transformer
• Node / JS