astralanx applies genetic programming to stock-picking: it generates candidate strategies, backtests them, keeps and recombines the best, and repeats this over many generations.
What a "strategy" is
A strategy here is neither a hand-written rule nor a neural network. Each one is a small abstract syntax tree built from a domain-specific language of financial primitives — indicators, comparisons, arithmetic, etc ...
How the search works
- Generate a population of random strategy trees.
- Backtest each one (see the methodology).
- Score on various metrics versus the rest of the population.
- Apply the GP algorithm.
- After many iterations, determine the best results
From engine to live site
A chosen strategy is deployed to the live dashboard: its formula produces target weights that rebalance every one to two months, and a simulated portfolio paper-trades those weights daily. The strategies whose formulas stay private show their performance and aggregate sector exposure only; a few fully open strategies show everything, end to end.