Crypto expert

guest@rkulskis.github.io:~/research$ echo LINKS: $(ls -a)
LINKS: . .. crypto_quant.org parallel_hdd.org scheduling_visualizer.org

About

Our conjecture is that no single trading algorithm fits every token swap pair due to the volume, volatility, and tribal knowledge specific to each security, so we want to train an expert that can choose a trading algorithm given a context which maximizes profit.

Training

Given:

  • \(pair\): of two coins \((c1, c2)\)
  • \(DB\): database of historical crypto trades
  • \(A\): set of stateless algorithms, each which requires their own context
  • \(C\): set of context state generators, each may be used by one or more algorithms and takes the same arguments, e.g.: \[window=(t_{start}, t_{end}), DB\]
    • can query external context on internet as necessary
      • idea: maybe host a version controlled thing for wikipedia which we use for seeing changes over that \(window\)

Output:

  • Simulation results of that trading strategy over time for that window
  • Q: should we output the context that was used? what about global context?

Use this output to train AI on whether that strategy works in that context.

Runtime