The Mill Run #3: 43.6% before costs, a broken pairs trade, and the model behind the number

A backtest can earn your attention before it earns your capital. This issue starts with an options model reporting 43.6% annualized mean return before costs. Another study follows two familiar stocks into a trade that never reaches its exit. The useful part is what happens when each headline meets the assumptions behind it.

These are research reviews, not completed Finance Paper Mill replications. All performance figures below come from the authors' backtests. My contribution here is source checking, interpretation and a proposed next test. A research priority is not a trading recommendation.

A thin week for actionable research. Most of the scored intake was code repositories. The four selections below received two Watch verdicts and two Reject verdicts from the mill. I kept the rejected cases for their failure lessons, not as trade candidates.

The intake, September 26 through the October 1 morning run: 392 records entered the mill and 28 have recorded scores. This issue selects three distinct academic papers from that pool, plus one August backlog revisit. None appeared in Mill Run #1 or #2.

1. An options model with a 43.6% gross return, and a cost curve to read first

Wee Ling Tan, Stephen Roberts and Stefan Zohren, Taming the Greeks: Option Portfolios with Inductive Biases.

Gross: 43.6% annualized mean return and a 2.339 Sharpe for the reported DP (L1) model. This is not a CAGR. The study uses Nasdaq 100 options from 2010 to 2023: next-month straddles, adjusted daily. Penalizing directional exposure raises Sharpe from the unpenalized model's 2.131.

Costs: Sharpe falls to 1.574 at 50 basis points and 0.860 at 100. These are modeled charges per unit of turnover, not measured fills. The table reports selected best-performing variants, not an executable break-even spread.

My read: Watch, with execution still unproven. I would freeze one specification, verify contract availability and quote timing, and apply executable spreads, financing and margin constraints to test whether the improvement survives real trades.

2. From the backlog: training for wealth growth still does not make the best portfolio

Jiayu Li, KellyBoost: Growth-Optimal Portfolio Construction with Gradient-Boosted Trees, August 2026.

Gross: 6.2% annualized geometric return for KellyBoost versus 8.2% for a two-stage forecast-and-optimize model. The comparison covers January 2013 to July 2026 and 163 monthly decisions. KellyBoost learns portfolio weights directly from a wealth-growth objective. Its reported maximum drawdown is 46.2%.

Costs: at 20 basis points one-way, KellyBoost returns 4.5% annualized and the two-stage model returns 4.7%. On those rounded figures, the two-stage model's 2-percentage-point gross lead shrinks to 0.2 points after costs. This is the paper's turnover-based sensitivity, not a universal cost estimate.

My read: Watch; test sizing before adding complexity. Every pairwise confidence interval in the main results table includes zero. The author also discloses that disappointing evaluation results prompted another design iteration. A fresh test should freeze the design and compare conservative allocations at the same risk and cost budget. I revisit this August paper, previously covered on the site and in the September 29 Notes pack, for the cost comparison.

3. Pepsi versus Coke: the spread that did not come back

Davide Graziano, From Cointegration to Out-of-Sample Failure: A Pairs-Trading Case Study on PEP/KO.

Net in sample: a 0.67 Sharpe. Remove the COVID window and it falls to 0.28. The author estimates the relationship on 2013-2018 data and optimizes trading thresholds over 2018-2023. The baseline charges 20 basis points on entry and exit of each leg.

The subsequent frozen-parameter test loses $28,131 on $100,000 initial capital, with just one trade. The exit condition never triggers and the position is liquidated at the sample end. The paper's September 2026 version labels that evaluation period as beginning in 2023, without a precise endpoint in its results table. One trade cannot support a reliable Sharpe estimate or a verdict on all pairs trading.

My read: Reject this implementation; keep the failure lesson. A historical statistical relationship is not an exit guarantee. I would require an observable rule for retiring a pair and test it across many pairs on untouched data before reconsidering. Lower costs alone cannot establish that the relationship will recover.

4. A 1.03 Sharpe from reading the news, with information the live trader would not have

Kevin Foley, Jonathan Hartadi, Shivesh Prakash and Swapnil Vatsal, From Word Counts to Context: Topic Models for Asset Pricing.

Reported: a 1.0334 excess-return Sharpe for the strongest exploratory specification. The financial evaluation spans 41 months, September 2020 to January 2024. The raw portfolio's maximum drawdown is 62.23%.

The authors replace word-count topics with contextual text representations and extend the portfolio construction. But topics and some scaling use the full sample, and the extension was chosen after inspecting the evaluation period. The Sharpe improvement is not statistically established.

Costs: no verified trading hurdle from this result. The paper distinguishes its rescaled display charts from a prospectively risk-managed, cost-adjusted strategy. Treating the headline as a net trading result would skip the central problem.

My read: Reject as evidence of a tradable edge. Reconsideration would require a frozen text model, training-only scaling, news available before each decision, and an untouched evaluation period. The same study arrived through arXiv and SSRN; it counts once here.

What deserves the next test

The options paper is the Watch idea I would examine first: it reports a risk-control comparison and a cost sweep, but executable returns remain unverified. KellyBoost stays on Watch too. Its attraction here is the cost comparison, not evidence that another model should run a reader's portfolio.

I am not moving either rejected paper onto a replication schedule. PEP/KO illustrates a relationship that failed to persist; the topic-model result needs a clean information timeline before it can support an investment case. Those are reasons to retain the lessons, not promote the strategies.

Read Mill Run #2 for the previous selection, or browse the research archive for published reviews and house tests.

Subscribe to Finance Paper Mill for the free Mill Run and research previews. Paid articles go deeper into completed house tests, execution assumptions and the limits of the evidence.

I run Finance Paper Mill and curate its research, drawing on 30 years of hedge fund experience. AI assists discovery and drafting; I select the research and approve publication. Research commentary, not personalized investment advice.