Six items this week, and most of them turn on the same question: can the reasoning behind a decision be reconstructed by someone other than the system that produced it? An investment output can be accurate, fluent, well documented, and still be unreconstructible. For someone building a long-term portfolio, the method lesson is that a result you can trace and a result you can only accept are two different assets, even when the numbers are identical.
A single trade is a harder test than a track record. The CFA Institute asks the question that reveals whether a quant manager can explain a single trade: reconstruct one specific position from the portfolio. The article separates two things that are easy to conflate — attribution, which identifies the signals that contributed to a decision, and explanation, which connects those signals to economic logic and says why they should have produced that outcome. It also warns that post-hoc explanation tools can manufacture confidence by appearing rigorous without faithfully representing what the model computed. Aggregate performance cannot distinguish the two, which is exactly why the single-trade version of the question does work the track record cannot.
Delegating the thinking is a governance problem before it is a technology problem. The CFA Institute’s piece on the risks of cognitive delegation and accountability describes professionals leaning on generated output before forming their own understanding of the analysis underneath it. The authors’ concern is structural rather than technological: responsibility does not transfer to a machine, so a process that quietly relocates judgment ends up with decisions nobody can attribute and conviction resting on outputs nobody has examined. Their proposed direction is to design for friction — clear decision rights, documentation, and audit mechanisms — so that the human judgment stays where the accountability already sits.
A benchmark for machine agents that scores work, not prose. FinSkillBench proposes an evaluation suite for whether language model agents can actually perform investment management tasks. What makes it interesting is the definition of success. The authors argue such a system must retrieve point-in-time data, assemble correct computational inputs, invoke specialized methods, and produce auditable structured outputs — none of which is demonstrated by generating plausible text. That list happens to be a decent description of what any research process owes a reader, machine or otherwise.
A well-known regularity moved when the measuring rule moved. More Frequent Than You Think: Revisiting Capital Structure Adjustment takes a documented finding — that companies adjust their leverage infrequently — and shows it is sensitive to two methodological choices: high thresholds for what counts as an adjustment, and reliance on net balance-sheet changes. Using lower thresholds and gross flows from cash-flow statements, the authors report adjustment is far more frequent than previously documented, with pronounced differences by company size. The subject is corporate finance, but the lesson travels: an empirical fact is a joint product of the data and the rule used to read it, and the rule is usually the part nobody restates.
Re-running your own published claim is itself the finding. Alpha Architect has published a follow-up to a 2017 study of its own, revisiting whether volatility information improves a trend-following allocation model with nearly a decade of out-of-sample evidence. The original idea was that volatility may carry information about how quickly momentum ought to be measured — a shorter window when conditions are turbulent, a longer one when they are calm. The out-of-sample years are the part that matters methodologically: they could not have influenced how the rule was built. Readers who want the results will find them at the source; the practice of publicly re-testing a claim you published yourself is the transferable habit.
Deciding the risk budget first makes the later decisions accountable to something. A revised working paper on dynamic tracking error and the total portfolio approach argues that the difference between two institutional frameworks reduces to one variable: how much tracking error a board grants. In the authors’ framing the first decision belongs to the board and is the drawdown it can tolerate, with the benchmark and the tracking-error budget following from it rather than preceding it. Active risk is then spent against that budget and reduced as the fund approaches its limit. Set up that way, every later decision has a stated constraint to be explained against — the same ordering the treasury view of risk starts from, arriving here from institutional portfolio management.
The throughline. Reconstructibility is the thread. A trade you can rebuild from named signals, a decision whose owner is identifiable, an output that arrives with its inputs auditable, a statistic reported alongside the threshold that produced it, a rule re-run on years it never saw, a budget set before the positions it governs — each of these is the same discipline in a different setting. Bank treasuries institutionalize it in an unglamorous way, by refusing to let a number size a position until someone else can reproduce it. Applied to a private portfolio it produces something more durable than a good year, which is why the portfolio is not the product and the ability to explain it is.
Curious where your portfolio’s risk structure stands? The free ETF Portfolio IQ Score is one way to see.