Websocket stream spot, chronological order of data?

[CharlyEmpereurmot]

Hi! So I have been using this package for a while, thanks again for this Oliver :)

I'm applying some machine learning models to predict the next price action for a given period, just UP or DOWN. My models are calibrated according to historical data, then I stream (spot) data from Binance and make predictions, act accordingly. The predictions yield a continuous value within [0, 1] and I have a hard threshold that says for example if prediction > 0.55, then next_price_action = UP.

What is puzzling me is that I see some discrepencies in the predictions my models are making, when it's using historical data VS the live streamed data. Sometimes I will get prediction = 0.56 using historical data and from the live streamed data I will get prediction = 0.54. After investigating quite a lot, I don't think the problem comes from my code, but rather from the underlying data the models are fed. I'm NOT using a callback, but rather stream_payload = binance_websocket_api_manager.pop_stream_data_from_stream_buffer(). And I trigger a new predicton as soon as I see a new timestamp for a candle opening, on a given pair I want to trade.

So I'm thinking now, the problem is that Binance does NOT send the candle information in the exact chronological order and/or these data do NOT come to my server in the correct chronological order, for whatever else reason.

Have you guys experienced similar behavior of the websocket streamed data? I have seen around that this is to be expected, but I did not think this would be so frequent. I had understood this could happen only "under Binance servers heavy load periods". What are you guys doing to handle this kind of problem?

Thank you very much for any advices you may provide!

[oliver-zehentleitner]

Hi Charly!

if you use historical data, you have only full candle sticks. (the closing one of each interval you used)

if you subscribe to websocket you get an opening candle with 0 vol i think and then you receive updates if there are trades. so this is different from trainging data.

you need to record this data to a database and train your network to it if it should be able to use data between two closing candles.

the last kline should have is_closed = True

Between 2 closing candles is minimum 1 minute, for me i think this never happend to be in the wrong order....

did you try to log your received candles and controll it?

[MoeQatoum]

Welc. Yeah I guess it's binance internal issue and nowadays is very nice time to stress test your code the market is very volatile.

Please let us know how is ai doing with guessing price direction. I always thought it's impossible for ai to predict price direction as it's completely random, so never tried it. Please let us know the results.

[oliver-zehentleitner]

A very good resource to this topic: https://github.com/PacktPublishing/Machine-Learning-for-Algorithmic-Trading-Second-Edition

[CharlyEmpereurmot]

It's not an easy task, but there definitely are ways. Even in randomness (or close to it, anyways) you can isolate some signals, but you need to find a way to select the strongest ones among the noise. Then forget about 70% accuracy/precision, this cannot happen (or very rare cases, not stable long-term). But you can already do stuff with > 53% for trading.

[oliver-zehentleitner]

I think the key is the training data :) all depends from good classifications

What good or bad signals are is the task of the learning algorithm.

[CharlyEmpereurmot]

Yes, and of the metrics used! What do you mean exactly about the training data? Something like using 3 classes (having a neutral one) instead of binary classifying? Maybe also using stratification for the cross-validation? I would love to hear a few more words about that :)