Finding a pattern and keeping the value it seems to promise are two separate problems. Much of this week’s research sits in the space between them: signals that are statistically real but too small to survive trading costs, rules for combining signals that do not deliver what they appear to, and tools whose output shifts with the wording of the question. The lesson for a long-term investor is one of method: an approach is measured by what it keeps after frictions, not by what it detects.
Predictable is not the same as exploitable. A CFA Institute piece, Can AI Agents Beat the Random Walk? Not So Fast, summarises published research by Sami Küçükoğlu that tested whether deep reinforcement learning agents can profit from market patterns that statistics can detect. Once transaction costs were included, the agents frequently failed to exploit long-memory dynamics. Some stopped trading altogether, and others settled into losing behaviour. The author calls the dividing line a “learnability threshold”. The practical framing offered is that these tools are more useful for stress-testing an approach than as a source of return.
A falsification study where nothing passed. A revised version of Structural Limits of OHLCV-Based Intraday Momentum Signals (first posted in May, updated this week) tested fourteen families of intraday momentum signals on an equity-index futures contract, using 947 trading days of five-minute data from 2021 to 2025 and walk-forward validation. Each family had to clear five criteria, including statistical significance out of sample and a positive result after realistic execution costs. None cleared all five. Eleven failed on costs alone, because their gross gain per trade sat below the friction threshold. What makes the study useful is its design: it also ran two signals built to pass, and both did, which shows the test could detect an edge had one existed. A test that cannot fail tells you nothing. This one could, and it did.
Capping correlation between signals does not settle what their combination does. Separated Signal Libraries: Packing, Saturation, and Joint Spectral Limits, posted this week, studies what happens when signals are admitted to a library only if they are not too correlated with those already in it. The paper shows that this separation rule, by itself, guarantees little about how the combined library behaves. Depending on the admission rule and the pool of candidates, the equally weighted sum of a very large library can line up with its dominant common component or sit orthogonal to it. The results are theoretical and validated numerically, without market data. The point carries over to fund selection: a pairwise correlation limit is a filter, not a description of the portfolio it produces.
How a fund’s tax structure works, and where the rules draw the line. A research post from Alpha Architect, 351 ETFs: Tax-Free “Diversification” Is Supposed to Hurt, explains two provisions that operate together. Section 351 allows securities to be transferred into a newly formed ETF without tax at the time, carrying the old cost basis into the new shares. Section 852(b)(6) lets a fund deliver appreciated securities to redeeming shareholders in kind without recognising a gain, which is the mechanism behind much of the tax efficiency of ETFs. The authors trace roughly 60 years of rulemaking and find a single consistent concern: preventing a concentrated position from becoming a diversified one without tax. When a transfer diversifies the contributor’s holdings, the contribution becomes a taxable exchange. The post follows a July 2026 Treasury comment on these transfers, and it is written as general education, not as tax guidance.
The wording of a question changes the answer. The CFA Institute report Managing LLM Bias in Investing tested 900 prompts across nine language models in ten investment scenarios. Presenting equivalent information as a gain rather than a loss moved the models’ judgements, most strongly when evaluating drawdowns and client retention. Rewriting prompts to present both the positive and the negative view produced the largest reduction in that framing bias. A separate bias-detection model helped less, and the authors conclude that human review remains the most effective control. The same bias is well documented in people, so the finding is less about machines than about how any analysis depends on its framing.
For readers closer to drawing an income. An Alpha Architect summary of Social Security’s role in federal debt covers a 2026 working paper by Romina Boccia and Ivane Nachkebia. It argues that the program should be judged by its annual cash flows as well as by its trust fund. By the authors’ calculations, cash-flow shortfalls since 2010, together with the interest on them, added more than $1.5 trillion to federal debt between 2010 and 2025. For planning purposes, the useful distinction is between scheduled benefits and benefits payable under current financing. A projected benefit is itself an assumption that depends on future policy, so it can be examined as a range rather than taken as a fixed number.
The throughline. Every item this week separates a signal from what survives of it. Costs erase real patterns, and a correlation cap leaves the combined library undetermined. Tax rules turn on what a transfer actually does, framing moves a model’s judgement, and a scheduled benefit rests on policy. The durable habit is to ask what remains after the frictions, the structure and the assumptions, which is the same question risk-budgeting asks of every position, as discussed in what a bank treasury knows about risk.
Curious where your portfolio’s risk structure stands? The free ETF Portfolio IQ Score is one way to see.