Pulse — the research engine

Pulse is the quantitative research module inside Atlas: 18 years of exchange-sourced market data, an engineered feature library, and a validation pipeline every model must pass before deployment. Negative results are documented with the same rigor as positive ones.

Research approach

Feature engineering over model complexity

Research effort concentrates on informative representations of market state — price structure, volatility regime, microstructure, time-of-day, trend state — each validated for incremental predictive value and monitored for decay as market conditions change. Admission to the library is itself gated: every candidate feature is tested against calibrated noise floors before use, and rejections are recorded alongside admissions. Model architectures are commodity components; the feature library and the validation infrastructure are the durable assets.

Validation

Candidate models are evaluated walk-forward on held-out data, against permutation-based significance floors, with explicit multiple-testing control. The process is symmetric about outcomes: when short-horizon direction prediction was tested across two feature generations and horizons from three to forty minutes, no statistically significant edge was found — the program was closed and the result documented. The model class that passed validation answers a narrower question: the probability of significant near-term movement, independent of direction.

Deployed model

Volatility state

The first production model estimates the probability of a significant index move within the next fifteen minutes. It was trained on 18 years of data and evaluated on a held-out period excluded from all training and tuning.

18 yrs
Training tape
757k
Bars in evidence
0.79
Held-out AUC
3.7×
Precision vs. base rate, top-confidence band

On the Terminal, the model surfaces as a single tile: the current probability, its confidence level, and the validation claim it is graded against. Scores carry their timestamps; a stale score is displayed as stale. The model is advisory — it does not place, size, or block orders.

Model lifecycle

Deployment is monitored, promotion is evidence-gated

Every published score is recorded at serving time. Realized outcomes are evaluated nightly against the claims made at validation — realized precision versus validated precision — and the resulting scorecards accumulate in the desk's reports. Promotion beyond observe-only mode requires sufficient forward evidence, a requirement enforced by the platform rather than by convention. Models whose forward performance decays are retired through the same registry process that deployed them.

Design principle: staged rollout constrains a model's authority, not its quality. The best validated model serves — monitored, graded, and subordinate to pre-trade risk checks on every order.

Research program

Specialist models, shared governance

The architecture is designed for a portfolio of narrow, individually validated models — each answering one question, each monitored and replaceable — rather than a single general predictor. The registry, validation pipeline, and forward-grading machinery that govern the first model are built to govern many, across instruments and timeframes. The current expansion applies the same discipline beyond the intraday index tape: a daily cross-sectional panel spanning the full NSE equity universe, with delistings and symbol changes handled point-in-time.