pip install ccxt torch numpy
run: >conda activate gpt-gpu python .\index.py
prompts: 1. create a 8b neural network (ai) that will consume live and historical HLOCv (candle sticks) data with a specific time window and in different time periods (1s, 1m 15m, 1h, 1d) and perform buy/sell operations. It will be based on the latest RL unsupervised training techniques, will continiously and retrospectively improve itself (without entering separate modes for training/inference) and the info it can digest will be able to be extendable and dynamic. for example, we should be able to feed sentiment analysis on current X feeds or news. We will also prepare/ calculte various indicators on top of the incomming HLOCV data (stocastic, rsi, etc - all most popular). we should be able to support up to 100 indicators (additional data) channels. The signals of the NN will be used by a bot first to trade on Solana using jupiter api. 2. let's continue with the implementation beyond skeleton. we should be able to do the first run with real live BTCUSD data. let's use the best free sources for fin data. we can use mexc api for making market orders 3. how to prepare backtrack data and feed it to the NN to train it? let's create the training loop. we should buy low & sell high. we should be able to backtest for arbitrary period in the past 4. save best model after the training set and load it when new training session starts. keep last 10 and best 10 models stored 5. draw a chart after the training session. show buy/sell actions during training and PnL after each order close 6.
the NN does not perform any actions. also, we should get 1500 candles on 5 different timeframes - it should help us find the ups and downs and the technical context. use the volume information as well, calculate common indicators, and feed all that data to the model so it can make 'informed' action. also, test if the actions were successful and profitable. 7. our fittness/reward function should take into account the PnL on first place, and the amount of profit per price variation/volatility.
model should be incentivised to do trades. we should be able to work with historical data, but also with real-time data stream - adding to the cached candles 8. The Nn should have one task - to predict next low/high on the short term charts ( 1m, 5m or other - configurable) based on the all past info in parallel from all the different timeframes candles and all the passed indicators candles. It should also have a dedicated NN module todiscover and pay attention to spefic parts in the charts - building and training it's own indicator in a sense. We later use the predicted high/low 5m/1h in the future to buy now and sell later or to short now and close later in the bot. We will have a threshhold of certainty to act, and also do it only if multiple timeframes predictions align. So we may use a transformer module to predict future candles and train that with RL while the candles are rolling until the NN can predict with small loss 9. modify the chart so we can see if the buy/sell is in or out. also, draw a dotted line between the buy and sell points for each position
we're trying to create a 8b neural network (ai) that will consume live and historical HLOCv (candle sticks) data with a specific time window and in different time periods (1s, 1m 15m, 1h, 1d) and perform buy/sell operations. It will be based on the latest RL unsupervised training techniques, will continiously and retrospectively improve itself (without entering separate modes for training/inference) and the info it can digest will be able to be extendable and dynamic. for example, we should be able to feed sentiment analysis on current X feeds or news. We will also prepare/ calculte various indicators on top of the incomming HLOCV data (stocastic, rsi, etc - all most popular). we should be able to support up to 100 indicators (additional data) channels. The signals of the NN will be used by a bot first to trade on Solana using jupiter api. we're stuck, and the code needs fixing The Nn should have one task - to predict next low/high on the short term charts ( 1m, 5m or other - configurable) based on the all past info in parallel from all the different timeframes candles and all the passed indicators candles. It should also have a dedicated NN module todiscover and pay attention to spefic parts in the charts - building and training it's own indicator in a sense. We later use the predicted high/low 5m/1h in the future to buy now and sell later or to short now and close later in the bot. We will have a threshhold of certainty to act, and also do it only if multiple timeframes predictions align. So we may use a transformer module to predict future candles and train that with RL while the candles are rolling until the NN can predict with small loss.
make sure we feed the NN different timeframes in parallel in different "channels" and the indicators as well. we should somehow synchronize them so the NN can make sense of the data and know how 1m 5m and 1h data syncs. incidators will be calculated for each timeframe channel separately. additionaly there should be room for meta learned parameters and indicators. NN will produce predictions for the next couple of candles. that will be used to generate signals and also for backtesting with RL. we buy at the bottom of anticipated valley and sell on the top
existing (running but unfinished ) code:
implement these suggestions into our code and add arguments for easy switching of modes:
- train (only): pool latest data and use it for backtesting with RL to learn to detect peaks/valleys
- live: load best checkpoint and latest HLOCv data to actively generate trade signals, but calculate and back propagate errors when closing positions. optimize for profit in the reward function
- inference: optimize model loading for inference only - load historical data and periodically append new live data and generate siganls but without active RL