Does research explain momentum as investor psychology?

Not exclusively. A study summarized this week argues that a large share of momentum profits line up with institutional cash-management mechanics rather than sentiment: investors who need settled cash before month-end tend to sell their losing positions, which pushes those prices down in a repeating intramonth pattern. It is one competing explanation among several, and it illustrates that a pattern in returns can have a plumbing cause rather than a behavioral one.

Why has the diversification between stocks and bonds looked different recently?

Research discussed this week documents that the correlation between stocks and bonds has been higher over roughly the past five years than in the preceding decades, alongside weaker bond returns. The educational point is that diversification is a measured relationship over a chosen window, not a fixed property of an asset class. A number that was reliable in one period is a measurement, and measurements move.

Why is it hard to evaluate an AI system that forecasts asset prices?

A new working paper names two problems. Models are trained on all available data, so any historical backtest can be contaminated by look-ahead bias. Markets are also reflexive: as more investors adopt a system, prices adjust and the pattern it learned can erode. The authors propose a real-time, out-of-sample benchmark to sidestep both, which is a methodology contribution rather than a claim about returns.

This week in research: the machinery underneath the number

This week’s research keeps pointing at the same thing from four directions: the number you are looking at is downstream of machinery you usually cannot see. A momentum premium may be a settlement calendar. A diversification benefit may be a correlation window. An impressive artificial-intelligence backtest may be a look-ahead artifact. For a long-term investor, the lesson is about method, not about what to hold — before trusting a figure, it is worth asking what produced it.

Momentum may be plumbing rather than psychology. A study summarized in The Intramonth Momentum Cycle proposes that momentum profits are largely driven by institutional cash-management mechanics. Investors who need settled cash before month-end systematically sell their losers, and that recurring flow leaves a footprint inside the month. Momentum has been explained for three decades by investor psychology, delayed information diffusion, and risk compensation; this paper argues the calendar of cash settlement does a lot of the work. The method lesson is durable regardless of whether the argument survives replication: a persistent pattern in returns can have an operational cause, and knowing which cause you are relying on changes how much confidence the pattern deserves.

Diversification is a measurement, not a property. Research on the recent behavior of stock-bond correlation documents that over roughly the past five years, bonds have correlated with stocks more consistently than in the preceding decades, and paired that with weaker returns. The phrase “bonds diversify equities” is shorthand for a correlation estimated over some historical window. Change the window, and the estimate changes. This is one of the clearest examples of why a portfolio’s risk structure is worth re-measuring rather than remembered — the relationship that made a given allocation look balanced is itself a moving quantity, and a long-term framework has to accommodate that rather than assume it away. It is risk-budgeting showing up in the data, and it echoes what a bank treasury knows about risk: correlations are inputs you monitor, not constants you inherit.

A benchmark built to keep AI evaluation honest. In Assessing the Benefits of Optimized Agentic AI Systems for Asset Pricing, Ralph Koijen and Bradford Levy start from an evaluation problem rather than a result. Systems are trained on all available data, so historical analysis carries look-ahead bias; and markets are reflexive, meaning adoption itself erodes the patterns a system was trained on. Their contribution is a real-time, out-of-sample benchmark designed to sidestep both. The value here for an ordinary investor is the framing: when a tool is marketed on backtested performance, the first question is whether the test could have known the future, and the second is whether the edge survives other people finding it.

Tools carry preferences of their own. An audit of asset preferences inside financial language models asks whether large language models — now embedded in robo-advisors and trading agents — hold systematic built-in preferences toward particular instruments. The authors develop a three-level audit protocol, identify an internal representation with causal leverage over those preferences, and show it affects downstream allocation decisions. Stated as method: an automated adviser is not a neutral calculator. It has a disposition inherited from its training, that disposition is measurable, and measuring it is a transparency question rather than a technological curiosity.

Defaults keep outperforming intentions. How Do State “Auto-IRA” Policies Affect Household Balance Sheets? compares private-sector workers exposed to Oregon’s automatic-enrollment retirement policy with similar workers in states that had not yet adopted one. The authors find the policy associated with increases in retirement-account ownership and assets. The finding belongs to the same family as decades of behavioral research: the structure surrounding a decision moves outcomes more reliably than exhortation does. Readers closer to decumulation may find this the week’s most relevant item, since the same principle — structure beats willpower — is what a written withdrawal framework is for.

The throughline

Four unrelated papers, one shared shape. Momentum looks like sentiment until you find the settlement calendar. Bonds look like a diversifier until you re-estimate the correlation. An AI system looks skilled until you ask what it was allowed to see. Retirement saving looks like discipline until you notice who was automatically enrolled. In each case the visible number is an output of machinery, and the machinery is the part worth understanding. That is the whole argument for transparent, written, re-measurable method over a remembered rule of thumb: not because the rule is wrong, but because you cannot tell when it has stopped being right if you never look underneath it.

Curious where your portfolio’s risk structure stands? The free ETF Portfolio IQ Score is one way to see.

The throughline

Related