All articles
    quantitative-investing
    regime-aware-allocation
    self-learning
    agentic-research
    backtesting
    portfolio-allocation
    asset-allocation
    quantify
    market-regimes
    survivorship-bias

    Nine Experiments, One Result

    September 29, 2026redevops.io
    Nine Experiments, One Result

    We gave an agentic research loop one question: can market state improve portfolio allocation? Most early hypotheses did not survive. One did (dynamic equity sizing), and the failures are now evidence for what the system should explore next.

    We started with a simple question: does knowing what kind of market you are in help you invest in it?

    After a series of experiments, one result survived: market state appears useful for deciding how much equity risk to take.

    On a 345-month window from December 1997 through August 2026 (including the dot-com crash, the financial crisis, COVID and the 2022 rate shock) the dynamic strategy returned 7.62% a year, with a 0.91 Sharpe and −18.2% maximum drawdown.

    The best fixed allocation by risk-adjusted return, 20% equities / 80% investment-grade bonds, returned 5.17%, with a 1.09 Sharpe and −16.6% drawdown. A 100% equity portfolio returned 9.33%, but with a 0.66 Sharpe and −50.8% drawdown.

    The important comparison is dynamic versus the best constant allocation:

    +2.56%/yr (t = 2.24) over the full period and +5.15%/yr (t = 2.33) in the held-out confirmation period.

    The strategy does not beat equities on raw return. The evidence is that changing equity exposure with observable market state improved the return/risk trade-off relative to a fixed mix.

    The mechanism is visible

    periodavg equity budgetdynamic100% equity
    dot-com, 2000–0244%−3.5%−35.5%
    financial crisis, 2007–0925%−8.4%−41.4%
    COVID, 20203%+1.4%+3.3%
    rates, 202229%−10.6%−8.3%

    During the financial crisis the strategy reduced equity exposure to roughly 25%; during the COVID crash it fell to 3%. It also shows its limitations. In 2022 bonds declined alongside equities, so reducing equity exposure did not provide the expected protection.

    That failure points toward the next question: how much should be in equities, and where should the rest of the portfolio go?

    This is early self-learning, an unfinished strategy search

    The research process is closer to the way a system learns to play chess than to a traditional one-shot backtest.

    You do not teach it one supposedly optimal move and stop. It generates candidate decisions, plays them against alternatives, observes what happened, keeps useful patterns, discards weak ones, and uses the accumulated evidence to decide what to explore next.

    That is how we want the Quantify research loop to work.

    The early programme explored three broad levers: how much capital to allocate to equities, which stocks to hold inside the equity allocation, and when portfolio changes are worth executing.

    So far, equity sizing is the strongest result. The other branches should not be interpreted as exhausted.

    Individual-stock selection in particular has not had the same quality or depth of historical evidence. Our Norgate price history retains delisted securities, giving us a survivorship-aware price universe. But the earnings data used in some stock-selection experiments came from Yahoo and is heavily incomplete for securities that are no longer listed. Earnings coverage is 99.6% for current names but only 13.6% for delisted names.

    The earnings-dependent experiments also cannot reach as far back as the allocation test. That matters because extending the allocation experiment into the late 1990s added market environments that were largely absent from the original 2010–2026 window.

    The current result is narrower and is not "stock selection does not work."

    We have not yet demonstrated persistent stock-selection alpha with the data available to this programme.

    A proper test needs longer point-in-time fundamental/earnings history with coverage of delisted securities. Until then, stock selection remains an open, unresolved branch of the self-learning process.

    Why the longer history mattered

    Our original evaluation window began in 2010. On that period, equities dominated bonds so consistently that a dynamic equity-sizing strategy was difficult to distinguish from simply holding more equities.

    Across ten market-state buckets, the ex-post best equity budget landed at the maximum tested allocation in nine of them.

    The sizing experiment does not require earnings data, so we extended it back to December 1997. That introduced the dot-com unwind and the financial crisis and gave the hypothesis environments in which it could actually fail.

    It survived.

    This is also why we now treat matched simple controls as mandatory. A dynamic strategy should not get credit merely for carrying more average market exposure than its baseline.

    What the system learned

    The useful output is larger than one allocation rule:

    • Market-state sizing deserves continued use and measurement. It survived the longer historical window and held-out confirmation.
    • The destination of defensive capital remains unresolved. The 2022 result shows why a two-asset equity/bond formulation is incomplete. Gold, cash, commodities and other defensive or alternative exposures are natural candidates for further testing.
    • Stock composition remains underexplored. The available price data is strong, but the point-in-time earnings/fundamental history is not yet sufficient for the same long-window test.
    • Complexity has to beat simple controls. Dynamic rules are now compared with constant allocations carrying comparable exposure so beta is not mistaken for intelligence.

    The loop is simple:

    hypothesis → experiment → control → outcome → evidence → next experiment

    Some branches die. Some remain unresolved because the evidence is inadequate. A few earn more capital and more testing.

    What's live

    The current allocation is published at quantify.club/research under Agentic, together with the growth of $10,000 against fixed-allocation baselines and every allocation change since 1997.

    As of 31 August 2026, the target is 65% equities / 35% fixed income.

    The equity leg uses SPY total return and the non-equity leg uses the FTSE US Broad Investment-Grade Bond index. Signals are observed on or before each decision date, and the state-to-weight mapping is estimated only from prior months. The strategy rebalances monthly at month-end and charges 10 basis points on turnover.

    The next iterations give the research loop more decisions to test, without adding complexity for its own sake: where defensive capital should go, whether better historical data can uncover persistent security-selection effects, and whether those effects add value independently of overall equity exposure.

    The strategy is still learning.


    Historical backtest, no live track record or investment advice. Past performance does not indicate future results. The current research uses survivorship-aware historical price data, but some stock-selection experiments rely on materially less complete historical earnings coverage, particularly for delisted securities. Taxes are not modeled.

    Want ReDevOps in your Google Top Stories & AI results?Follow ReDevOps on Google