[beoberha]

Sorry if I’m just bad at reading comprehension, but what exactly is being modeled here? It sounds like the author is just getting price data not unlike a candlestick chart. What I don’t understand is if the trading is happening within the 5 minute windows he mentions or if model is finding patterns across timeframes?

[Frummy]

So I don’t know this stuff so I’ll stochastically parrot a reply here of my understanding of what I just read.

The ML model is trained on the price spread and volume data collected with the cronjob every 5 minutes and stored in a database, and the trades with their respective data, but on previously collected data from 63 days up to 14 days earlier and then validated on the remaining later 14 days of data, apparently to not have temporal leakage which I googled and means unintentional backflow of training data which screws up the prediction. So to answer your question the model finds patterns over multiple weeks of data, to rank the best trading choices in the present moment. Then the java component executes the trade in-game with the player character.

What I personally don’t understand is if this data collection and training is rolling continually, or if he let it be trained one shot and then it’s smart enough to remain profitable. Don’t entirely understand the 49 day training, 14 day validation thing either.

[partypete]

That’s right - I retrained the model every 7 days. These experiment results are from the most recent one-week period.

[partypete]

The training data is composed of the bot’s previously executed trades from the last 63 days joined on the price/volume data. So the model discovers patterns in the price/volume data that have historically lead to well-performing trades. The dependent variable is gold/second generated from a historical trade.

[beoberha]

Thanks! But how was the bot running before the training data existed? I think that’s what I was missing. The baseline?