Research digest
Every issue of our research digest, as it was sent. The daily update leads with the three papers our screen found most interesting; the weekly digest opens with the week in numbers — the research funnel, the backtests and how the AI Quant portfolio did.
Steward applies a multi-agent LLM architecture to US equity and ETF investing, with outcome-based memory, gated prompt updates and immutable risk controls. It reports forward paper-account trading results, recorded-input replays and historical backtests of its deterministic risk core; these are distinct evaluations, not evidence of independently established live-money alpha.
The paper derives dynamic allocations that balance expected logarithmic growth against Range Value-at-Risk, producing state-dependent terminal payoffs and replicating strategies. It is directly relevant to portfolio risk management, but its model-based numerical illustrations are not evidence of realized trading performance. A daily implementation is backtestable, although it cannot reproduce exact continuous-time replication.
The paper derives causal position-sizing filters for return forecasts under inventory risk and transient trading costs. Its results are theoretical, not a reported backtest. A daily US-equity implementation is feasible, but its assumed price-impact kernel cannot be validated from bar data alone.
The paper reports long-short option returns from momentum and reversal signals based on prior intraday and overnight option returns. We could not construct its half-day returns because the platform lacks intraday option quotes and Greeks.
The paper derives an alpha-aware execution horizon and a front-loaded schedule under concave impact and Hurst-scaled price risk. Its reported 10–17% cost saving comes from its execution-cost model, not measured trading performance.
The paper uses portfolio-performance attributions to adjust machine-learning statistical-arbitrage allocations without retraining. It reports improved out-of-sample net Sharpe ratios for US equity portfolios, but our shorter data history cannot reproduce its full 2006–2024 test.
The paper tests a daily, news-informed actor-critic policy for allocating among U.S. equities, with downside-risk and turnover penalties. Its reported performance is for its own small equity portfolios; our shorter news history would not reproduce its 11-year test.
The paper derives an optimal investment position for a mean-reverting risky asset when price observations are delayed. Its numerical results illustrate a theoretical model, not measured trading performance on market data.
The paper develops a reinforcement-learning policy for setting market-making quotes using Hawkes-filtered order-flow history. Its reported reward comparisons come from simulated market-making environments, not a backtest on real market data.
The paper uses portfolio-performance attributions to adjust machine-learning statistical-arbitrage allocations without retraining. It reports improved out-of-sample net Sharpe ratios for US equity portfolios, but our shorter data history cannot reproduce its full 2006–2024 test.
Craft predicts next-day stock direction from historical prices and reports an investment simulation with Sharpe-ratio and drawdown results. Its price-only method can be tested on US equities, but the platform cannot reproduce its non-US evaluations or an exact historical S&P 500 membership universe.
The paper tests supervised-learning and reinforcement-learning policies for closing SPX short straddles and reports out-of-sample trading performance. Its intraday option-quote and execution requirements exceed the platform's data coverage.