Modeling the Market
M.H. / NQ Research
Intraday market researchNQ · 8:30–15:00 CT

Modelingthe market.

Systematic research into Nasdaq futures, market structure, and intraday trading behavior. Seven rule-based models, each tested on data that played no part in choosing it.

16 Oct 2026 · one session, read seven waysScroll
00 / Record

    01 / Model atlas

    Seven ways to read a session

      8:30 CT
        02 / Model

        03 / ValidationHow the models are judged

        A backtest
        is a claim.

        Every number above comes from a backtest, and every choice made while looking at a backtest is a chance to fit noise. These are the tests the seven models were put through, the corrections for the number of ideas tried, and the results they did not like.

        All seven together · cumulative R, after costs

        Correlation of daily R

        Models rarely lose on the same day; the highest pairs are the Market Profile models, which can trigger on the same drive day.

        Out-of-sample checks · per modelincl. failures
        Correcting for the number of ideas tried

        If N worthless variants are tried, the best still shows a positive Sharpe ratio. Its expected size (Bailey & López de Prado, 2014) is

        E[max SR] ≈ √V · [ (1−γ)·Φ⁻¹(1−1/N) + γ·Φ⁻¹(1−1/(N·e)) ]

        about 2.5 standard deviations at 100 trials and 3.5 at 3,000. The deflated Sharpe ratio is the probability that a model's Sharpe ratio beats that bar, given the sample length, skew and kurtosis:

        DSR = Φ( (SR − SR₀)·√(T−1) / √(1 − γ₃·SR + (γ₄−1)/4·SR²) )

        Because the true count of ideas tried is uncertain, each model reports it at 100 and at 3,000 trials. The probability of backtest overfitting (Bailey et al., 2017) asks how often the in-sample winner ranks below the median out of sample.

        Checks computed from the trade recordsnone used to build the rules
        Worse fills · total R vs extra slippage

        Extra ticks per side on every trade, on top of the one already assumed, as a share of each model's total.

        Since the rules froze · 23 Oct 2026

        Checked after each close on Databento one-minute data. Four weeks prove nothing yet; the record is kept so it cannot be chosen later.

        What did not survive
          04 / MethodRead before the numbers

          Assumptions

          All performance on this page is hypothetical. Results to 23 October 2026 are backtests on historical one-minute data with simulated fills; they are not realised trading profits. Fills on fast days can be worse than one tick. Results since 23 October 2026 are a forward record kept under frozen rules, still simulated on each day's data rather than broker fills. Nothing here is investment advice.