Skip to content
Ownsight

Published tests

Research

How each model in Ownsight was tested. The rules for every test were written down and committed before any results were run, and the results are published whether they flatter us or not.

  1. 01

    Risk of a bad day

    What we tested
    Which estimate of a 1-in-100 bad day holds up, across 7 test portfolios.
    Method
    Backtests of plain history, volatility-scaled history, GARCH and extreme-value models.
    Period
    Daily prices to October 2026
    Result
    Volatility-scaled history is the one shown. Plain five-year history failed on 6 of 7 portfolios.
    How sure
    Formal coverage tests (Kupiec, Christoffersen).
    Limitations
    Gold was borderline. Bad days cluster in crises, so a single number can understate them.
  2. 02

    Model baskets

    What we tested
    Five published ways to build a diversified mix, at five risk levels.
    Method
    Walk-forward: rebuilt every four weeks using only data available at the time, with costs.
    Period
    May 2015 to October 2026
    Result
    Results per basket and level, next to a fixed mix, 60/40 and equal weights.
    How sure
    Bootstrap ranges and significance tests (see the next row).
    Limitations
    One historical path; no tax; a small set of funds; periods before 2015 not covered.
  3. 03

    Are the basket results luck?

    What we tested
    Whether the baskets' results survive tests for luck and overfitting.
    Method
    Stationary bootstrap, deflated Sharpe ratio, probability of backtest overfitting, nested tuning.
    Period
    May 2015 to October 2026
    Result
    Published settings kept. A basket is only marked as ahead of the fixed mix if the gap is significant.
    How sure
    Overfitting risk rises with the risk level (higher at levels 4 and 5).
    Limitations
    These tests measure the past only; they can't rule out a different future.
  4. 04

    Rebalancing on example portfolios

    What we tested
    Whether self-adjusting rebalancing beats simple rules, on 800 invented portfolios.
    Method
    Eight kinds of investor; holding, resetting, blending and two self-training methods.
    Period
    January 2019 to October 2026
    Result
    Moving towards the baskets cut falls; neither self-training method beat a simple 50/50 blend.
    How sure
    Bootstrap p-values across portfolios.
    Limitations
    Invented portfolios, not real people's; no tax or trading frictions beyond costs.
  5. 05

    Machine-learning portfolio model

    What we tested
    Whether a machine-learning model beats the simple blend on types and years it never saw.
    Method
    Gradient-boosted trees, tested leave-one-type-out from 2023.
    Period
    January 2023 to October 2026
    Result
    Better than doing nothing, but no better than the simple blend, so it isn't used.
    How sure
    p = 0.44 against the blend: no evidence of an edge.
    Limitations
    Example portfolios only; never applied to a real person's holdings.
  6. 06

    Ownsight's language model

    What we tested
    Whether fine-tuning a small open model makes explanations more accurate and safer.
    Method
    LoRA fine-tune of Qwen2.5-0.5B, compared with the base model on 128 held-out items.
    Period
    October 2026
    Result
    More accurate figures and fewer advice-language slips than the base model.
    How sure
    A small held-out set; mostly one explanation format.
    Limitations
    A prototype that isn't running on the site; answers here come from Gemini, checked the same way.

Past results are no guide to the future. None of these tests says what any person should hold.