Overview
The market-warning work asks whether observable market and macroeconomic information can produce useful warnings before major declines. It evolved through many report versions, with simple baselines, additional features, historical crisis checks, allocation policies, and corrections to the data.
The central difficulty is that a small number of famous crises can make a rule look persuasive after the fact. Testing whether a warning fires before those events is much easier than demonstrating a useful prospective system.
A model must beat a relevant baseline
One evaluation compared a seven-feature logistic model against a simpler price-based specification across four panels. The larger feature set performed worse on Brier score and log loss. The design used chronological label-maturity and embargo rules, and retained convergence and implementation checks.
More features did not automatically mean better forecasting. Some evaluation segments had no positive events, which also limits what can be learned from their apparent performance. Model-fitting success is different from predictive success.
Two famous crises are not enough
A historical warning rule identified both of two selected major crises. But 161 of 200 count-matched random controls and 82 of 200 circular controls also captured both in the evaluated setup. The comparison substantially weakens the headline “two out of two.”
This is a useful example of why a baseline must preserve the relevant opportunity to succeed. A historical rule can have many opportunities to issue a warning; counting only the crisis hits leaves out how easily a less informative process could achieve the same result.
Reducing false alarms creates other costs
A later iteration reduced false Friday alarms from 122 to 38 while retaining two crisis hits on reused history. That apparent improvement came with costs, including delayed warning in one episode and lost coverage of smaller selloffs.
Another housing-related variant reduced a count from 209 to 29 but lost a historical crisis and was treated as a failure. Repeated changes on the same episodes remain development on reused history. They cannot create an untouched validation set or calibrate event probabilities from one or two crises.
Data revisions can change the conclusion
One correction introduced official lending and volatility releases to fill a large missing-data gap. After the correction, the reported compound annual growth rate fell from 11.91% to 11.47%, a recent-period return measure fell from 105.90% to 83.39%, and maximum drawdown improved from 19.56% to 11.26%.
Those mixed changes were retained. A revised dataset can weaken the return story while improving a risk measure. The correct response is to describe both, not select whichever metric supports the preferred narrative. Availability at the time of a decision still requires separate attention from a dataset’s present-day completeness.
V45: an unfavorable policy comparison
The V45 research brief examines 30 cases across 1,206 closes from October 2021 through July 2026. In the reported comparison, the C2 policy returned 62.43%, versus 82.75% for holding SPY, with the same 24.42% maximum drawdown. It did not demonstrate the intended improvement.
The report explicitly describes the analysis as retrospective. It is not untouched out-of-sample evidence. Data-vintage, settlement, and implementation assumptions limit an interpretation as tradable performance. Regression tests and a rational-arithmetic oracle help check implementation; they do not validate forecasting ability.
Scope and interpretation
The project includes model experiments, rule comparisons, data corrections, and a report interface. A working report is not proof of a live forecasting product. Large counts of internal checks or parameter combinations are not large counts of independent market observations.
My work is best represented by the questions pursued, the comparisons made, and the conclusions that the evidence would not support. I am especially interested in evaluation methods that make it harder to mistake historical selection for a reliable decision rule.