BLISP

BLISP is a tool that allows systematic AI-based alpha research (and probably also production trading) safer, repeatable and easier to inspect. It is effectively a layer which funnels the stochastic creativity of LLMs into templated instructions. There are several ways for free text to express the same calculation: for instance computing returns over "half a year" or "126 trading days". The BLISP system ensures that the user knows what an LLM is actually expressing, reducing the risk of low interpretability or repetition.

The system also creates a fingerprint of each calculation, recording the settings, data dependencies, versions, parameters, etc. This fingerprint should allow anyone to be able to faithfully reproduce any alpha, safely reuse results, and distinguish between changes which are trivial, and changes which are meaningful.

The author reports that unjustified tool use fell from 23% to 10% in his experiment, albeit on a small number of experiments.

We believe that a system such as BLISP would be important for letting an LLM conduct alpha research in a controlled way, leveraging its creative stochasticity, but ensuring traceability and reproducibility.

One potential limitation is that very similar signals (e.g. one with 253 day rolling window and another with 252 days) will be given completely different fingerprints; one cannot use the fingerprints to compute signal 'distance', and therefore the effective number of bets may be overestimated during LLM based research.

blisp.ai