DATA · Market data pipeline

From raw futures observations to one auditable model input.

The forecasting model only sees a compact pre-open signal. Reaching that signal requires a data pipeline that retrieves the correct futures contract, respects fixed New York timestamps, checks data freshness and preserves a reproducible audit trail.

The data pipeline defines what information reaches the model.

Graham Says measures Nasdaq-100 futures between the previous US cash-market close and 09:15 New York time. The engineering job is to make that measurement the same way every session rather than relying on a manually selected price.

Five steps turn market data into the model input.

01Resolve contract

Select the intended Nasdaq-100 futures contract and handle the roll explicitly.

02Retrieve bars

Request the market observations needed around the two fixed endpoints.

03Align time

Interpret timestamps in America/New_York and preserve the fixed observation window.

04Validate

Check endpoint availability, freshness, price sanity and contract consistency.

05Derive signal

Convert the two endpoint observations into the pre-market log return.

Once the two observations are fixed, the model input is fixed too.

r_pre = ln(NQ_09:15 / NQ_prior_16:00)
The public site publishes the derived signal used by the forecasting experiment. Raw licensed market-data operations, credentials and vendor responses stay outside the public repository.

What the pipeline guards against.

Contract identityRoll-awareAvoid mixing observations from different futures contracts across the measurement window.
Time conventionNew York timeThe window is defined in New York market time, not the computer's local timezone.
Endpoint qualityValidatedThe pipeline checks that observations used to construct the signal are usable before forecasting continues.
Public boundaryDerived output onlyOnly derived research outputs are published. Raw vendor data and credentials stay private.

Reproducibility starts with the inputs.

A one-line regression can still produce an irreproducible result if the underlying futures observation was selected by hand, taken from the wrong contract, or measured at a different time. The market-data pipeline exists to make the input definition operational rather than aspirational.