improve training and model data

docs/dev/architecture.md

@@ -6,17 +6,18 @@ III. models we currently use (architecture is expandable with easy adaption to n
 - cnn price prediction model - uses calculated multilevel pivot points and historical price data to predict the next pivot point for each level.
 - DQN RL model outputs trade signals
 - transformer model outputs price prediction
-- COB RL model outputs trade signals - it is trained on cob (cached all COB data for period of time not just current order book. it should be a 2d matrix 1s aggregated ) and some indicators cummulative cob imbalance for different timeframes.
+- COB RL model outputs trade signals - it is trained on cob (cached all COB data for period of time not just current order book. it should be a 2d matrix 1s aggregated ) and some indicators cummulative cob imbalance for different timeframes. we get COB snapshots every couple hundred miliseconds and we cache and aggregate them to have a COB history. 1d matrix from the API to 2d amtrix as model inputs. as both raw ticks and 1s averaged.
 - decision model - it is trained on price prediction and trade signals to learn the effectiveness of the other models in contribute to succeessful prediction. outputs the final trade signal.
 
 
-
  IV. by default all models take full current data frames available in the orchestrator on inference as base data - different aspects of the data are updated at different rates. main data frame includes 5 price charts
 class UniversalDataAdapter:
  - 1s 1m 1h ETH charts and ETH and BTC ticks. orchestrator can use and extend the UniversalDataAdapter class to add new data sources and data types.
 - - cob models are different and they get fast realtime raw dob data ticks and should be agile to inference and procude outputs but yet able to learn.
 
-V. hardware. we use GPU if available for training and inference for optimised performance.
+V. Training and hardware. 
+ - we should load the models in a way that we do a back propagation and other model specificic training at realtime as training examples emerge from the realtime data we process. we will save only the best examples (the realtime data dumps we feed to the models) so we can cold start other models if we change the architecture. i
+ - we use GPU if available for training and inference for optimised performance.