The volatility filter I employ keeps only bars in the ≥ 75th percentile of the 20‑day rolling volatility distribution.
The underlying random generator is the Mulberry32 PRNG, seeded for full reproducibility across machines.
My design emphasizes four core principles: surfacing ambiguity instead of guessing, deterministic simulation for research rigor, strict epistemic separation of facts, interpretations, and conclusions, and transparent provenance badges so users can always tell whether a bullet came from their own wording, an AI inference, an assumption, or a missing parameter.
I built the Trading‑research‑assistant repository on GitHub (1 star) in September 2026 and last updated it in November 2026. It is a TypeScript project that implements an AI‑native, five‑stage research workflow for quantitative trading ideas.
The assistant enforces a strict ASK → CLARIFY → DEFINE → TEST → LEARN lifecycle to avoid fabricated rules and hidden biases.
During the TEST stage it runs a deterministic backtest on a synthetic market generator seeded to produce exactly 1,250 NIFTY‑50 trading bars.
Results are presented in three epistemic layers: raw data facts, interpretive commentary, and cautious conclusions with risk disclaimers.
I designed the CLARIFY stage to surface latent assumptions—e.g., defining a "sharp fall" as a 1 % drop—so users must confirm parameters before any calculation runs.
Open questions include how well the synthetic generator captures real‑world dynamics, the optimal granularity for ambiguity detection, live‑data integration while preserving reproducibility, and UI designs for the CLARIFY‑DEFINE loop.
AI‑Native Trading Research Assistant Turn ambiguous natural-language market questions into structured, testable quantitative experiments with rigorous separation between data and AI interpretation.