AI-automatedresearch

AI is now making real discoveries. It disproved an 80-year-old Erdős conjecture1 and beat Strassen's matrix multiplication algorithm for the first time in 56 years.2

If AI can do research at this level, it should be able to find an edge in the markets too.

Can AI automate quant research?

To find out, we built a non-stop research loop. Autonomous AI reads past results, proposes a change to a strategy, backtests it on decades of market data after fees and slippage, and checks whether the result improved. If it did, the change is kept. If not, it's discarded, and the AI uses what it learned to plan the next experiment. It's a closed loop, fully automated, so each round builds on the last. The output is a log of experiments and, hopefully, a better strategy.

Market data is adjusted for fees, spread, slippage, margin, and liquidation assumptions. AI research uses chronological training and walk-forward validation periods. A final holdout remains untouched until a candidate is frozen, then a human gate determines whether it advances to live algorithmic trading.

Swipe to explore the research system

Most AI trading projects let a language model make judgment calls on each trade.3 AlphaStone uses AI to develop systematic strategies, and then trades them systematically.

Since 2021, AlphaStone has built its data pipeline, cost model, backtester and validation rules by hand. In 2026, while we were improving Model 5.5, AI began doing useful research inside that framework, generating ideas, testing them and building on its own results.

Markets are a hard place for this. Returns are noisy, conditions shift,4 and with enough tuning almost anything looks good on paper.56 So an improvement counts only if it holds up on data the AI never touched, and a human decides what moves forward.

Strategy philosophy

Multi-strategy

A single strategy rarely outperforms the market consistently over long periods, so AlphaStone combines multiple strategies to cover different market conditions.7

Absolute alpha

AlphaStone aims for absolute alpha. The goal is to outperform the market, not to be a passive investment.

Consistency in profits

AlphaStone favours strategies that perform steadily across market conditions, with extra care taken to keep drawdowns low.

The work so far.

  1. 2021FoundationAlphaStone began as a personal project to develop a systematic approach that could beat the market.
    • June 3AlphaStone Technologies SP was founded by Thomas Lee.
  2. 2022First live trading systemThe project moved from research into live testing. Model Ares and R&D framework V1 exposed the practical problems that do not appear in a backtest.
    • January 7Development of Model Ares began.
    • August 20R&D framework V1 was introduced and Model Ares entered live testing.
  3. 2023Framework rebuiltNew model generations and a rebuilt R&D framework expanded the research system, while failed experiments narrowed what was worth pursuing.
    • April 13Development began on Models 1 through 4.
    • June 18R&D framework V2 began development from scratch.
    • August 3Development of Model Ares was paused as research moved toward newer model generations.
    • December 16Development of Model 5 began.
  4. 2024Flagship Model 5 lineupModels 4 and 5 reached major development milestones. Model 5.2 became the working Model 5 line, and R&D framework V3 strengthened the foundation underneath it.
    • March 29Major development of Model 4 was completed.
    • July 25Major development of Model 5 was completed.
    • July 26Development of Model 5.2 began as the next stage of the Model 5 research line.
    • August 25Major development of Model 5.2 was completed. From this point forward, the name Model 5 refers to Model 5.2.
    • September 4Major development of R&D framework V3 was completed.
  5. 2025Improved framework V4R&D framework V4 and Models 5.3 through 5.4 were developed, followed by the start of Model 5.5.
    • April 15Development began on R&D framework V4 and Models 5.3 through 5.4.
    • July 20Development of R&D framework V4 and Models 5.3 through 5.4 was completed. Work then began on Model 5.5.
  6. 2026AI-assisted researchModel 5.5 was completed, a Model 5.6 attempt failed to produce a version worth advancing, and the project returned to improving Model 5.5 as AI became useful for research.
    • April 4Major development of Model 5.5 was completed, and development of Model 5.6 began.
    • July 26The Model 5.6 attempt did not produce a result worth advancing.
    • July 27Research returned to improving Model 5.5. This became a turning point as AI became materially useful for proposing, testing, and iterating on research ideas inside the existing framework.

The research loop.

AI proposes and tests candidate changes. The ones that improve on validation go to human review, then to a final holdout the AI never sees. Approved strategies are run by live algorithms. The roles are deliberately separated.

  1. AI research

    Candidate signals and strategy changes are proposed, tested and refined within the existing research framework. The results of each attempt shape the next one.

  2. Human review

    Human validation evaluates candidates out of sample, reviews walk-forward behavior and costs, and determines which advance.

  3. Live algorithmic trading

    Approved strategies are deployed to live algorithmic execution. The algorithm follows the approved rules, and individual trades are not overridden by discretionary judgment.

Human-built infrastructure

  • Data and market-data processing
  • Backtest and risk infrastructure
  • Out-of-sample and walk-forward validation
  • Live execution stack (Ares)
  • Reproducible research records

AI-assisted research

  • Proposes candidate signals and strategies
  • Explores multi-strategy combinations
  • Tests consistency across market conditions
  • Investigates absolute alpha hypotheses
  • Analyzes failed candidates
  • Suggests the next research attempt

Human validation

  • Maintains the validation framework
  • Checks for data contamination, look-ahead bias, and overfitting
  • Reviews out-of-sample results
  • Decides what advances

Current research uses a single AI model. Model attribution is provided for transparency. Comparative model testing will be reported when conducted under the same research framework.

Data preparation and integrity.

Every candidate uses cleaned market data, realistic execution assumptions, and a reproducible test configuration.

Data sources

Market data is cleaned, adjusted for corporate actions, and checked for survivorship bias before a candidate is tested.8

Cost modeling

Commissions, spread, slippage, market impact, financing, margin, and liquidation are modeled where they apply. Results are evaluated after estimated costs.9

Reproducibility

The same data, code, and parameters can be used to reproduce a test. This keeps comparisons consistent across research attempts.

Validation criteria.

Candidates must demonstrate out-of-sample performance, survive walk-forward analysis,10 remain viable under realistic cost assumptions, and be examined across market conditions before deployment is considered.

Fixed research periods

Complete calendar months are reported separately as training (Apr 2003–Dec 2018), validation (Jan 2019–Dec 2023), and holdout (Jan 2024–Jun 2026).

Validation-led selection

Candidates are ranked by validation average monthly NAV return after training review and strict accounting and risk checks. Holdout is displayed only after selection and never used to choose a record.611

Freeze before holdout

A candidate, configuration, and the exact data and code versions are recorded before the holdout is evaluated. The holdout is reported once; a failure is recorded rather than retuned away.12

Out-of-sample testing

The candidate is tested on data excluded from development. Data access and test timing are reviewed to avoid contamination and look-ahead bias.13

Cost and slippage stress

The result is recalculated with conservative trading-cost assumptions. Candidates that depend on optimistic execution do not advance.

Regime robustness

The candidate is examined across different market conditions. A result confined to one regime is treated as regime-dependent.4

References.

  1. 1

    OpenAI (2026). An OpenAI model has disproved a central conjecture in discrete geometry. OpenAI research announcement, 20 May 2026. openai.com↩

  2. 2

    Novikov, A., et al. (2025). AlphaEvolve: A coding agent for scientific and algorithmic discovery. Google DeepMind white paper, arXiv preprint 2506.13131. arxiv.org/abs/2506.13131↩

  3. 3

    Xiao, Y., Sun, E., Luo, D., & Wang, W. (2024). TradingAgents: Multi-Agents LLM Financial Trading Framework. arXiv preprint 2412.20138. arxiv.org/abs/2412.20138↩

  4. 4

    Lo, A. W. (2004). The Adaptive Markets Hypothesis: Market Efficiency from an Evolutionary Perspective. Journal of Portfolio Management, 30(5), 15⁠–⁠29. doi.org/10.3905/jpm.2004.442611↩

  5. 5

    Bailey, D. H., Borwein, J. M., López de Prado, M., & Zhu, Q. J. (2017). The Probability of Backtest Overfitting. Journal of Computational Finance, 20(4). doi.org/10.21314/JCF.2016.322↩

  6. 6

    Harvey, C. R., Liu, Y., & Zhu, H. (2016). … and the Cross-Section of Expected Returns. Review of Financial Studies, 29(1), 5⁠–⁠68. doi.org/10.1093/rfs/hhv059↩

  7. 7

    Markowitz, H. (1952). Portfolio Selection. Journal of Finance, 7(1), 77⁠–⁠91. doi.org/10.1111/j.1540-6261.1952.tb01525.x↩

  8. 8

    Elton, E. J., Gruber, M. J., & Blake, C. R. (1996). Survivor Bias and Mutual Fund Performance. Review of Financial Studies, 9(4), 1097⁠–⁠1120. doi.org/10.1093/rfs/9.4.1097↩

  9. 9

    Frazzini, A., Israel, R., & Moskowitz, T. J. (2018). Trading Costs. Working paper, SSRN 3229719. doi.org/10.2139/ssrn.3229719↩

  10. 10

    Pardo, R. (2008). The Evaluation and Optimization of Trading Strategies. 2nd edition, Wiley. wiley.com↩

  11. 11

    White, H. (2000). A Reality Check for Data Snooping. Econometrica, 68(5), 1097⁠–⁠1126. doi.org/10.1111/1468-0262.00152↩

  12. 12

    Bailey, D. H., & López de Prado, M. (2014). The Deflated Sharpe Ratio: Correcting for Selection Bias, Backtest Overfitting, and Non-Normality. Journal of Portfolio Management, 40(5), 94⁠–⁠107. doi.org/10.3905/jpm.2014.40.5.094↩

  13. 13

    López de Prado, M. (2018). Advances in Financial Machine Learning. Wiley. wiley.com↩

Research record.

The research log documents candidate ideas, tests, failures, and results as the project develops.

Research Live results