Create the exchange analytics server

[spencerbug]

This server receives the broadcast messages of updated individual positions of orders (filled/unfilled updates from the exchange server) It then creates all features and metrics, preprocessed, that can be passed to all the trading agents directly, so they can all share the same data. Data consists of

• A moving timeline of limit orders, normalized. This can be implemented with a circular buffer, copying the previous state each timestep, and then updating states as the updates are received from the exchange. The price ranges and the volume have to be z-score normalized to the rolling window, so they can fit into a finite sized circular buffer, For example: The price ranges from 2250 and 2350 for the last 300 timesteps. 2250 to 2350 would z-score map to -1.0 ... +1.0. Then we can arrange those into 10 buckets grouping together the average volume of those limit orders for each bucket. The volume of quantities for each bucket would also have to be z-score normalized.
• LSTM autoencoding of the moving limit orders
• Other metrics, z score normalized of their percent changes, over multiple windows of the ringbuffer
  • Price
  • SMA: Simple Moving Average
  • RSI: Relative Strength Index
  • ADX: Average Directional Movement Index
  • MACD: Moving average convergence divergence
  • WVAP: Weighted volume average price
  • CCI: Commodity channel index
  • TSI: True Strength Indicator
  • OBV: On-Balance Volume

All this data will be processed in realtime by a dedicated core or GPU, and the LSTM autoencoder will have its weights updated with online learning every few timesteps. The data shall be made available for every agent in the network, by using shared numpy buffers, or via UDP broadcast.

[spencerbug]

Updating this ticket, we will not be z-score normalizing the output, instead leaving that to the individual trading bots to do. Current analytics sever plan will output a ringbuffer containing: SMA, RSI, ADX, MACD, WVAP, CCI, OBV, TSI, Open, High, Low, Close, Volume, a fixed size depth chart with a variably sized price buckets using a pattern, and the full volume depth chart without bounds, this last one would be convolved for a variable sized input on the RL Agent