A prediction is only useful if you can hold it.
Fast-moving inputs can make a model rebuild the portfolio every month. The paper tests inputs, model sensitivity, and update rules in separate experiments. Open each measured result to see where implementation churn enters.
Feature speed makes an implementation problem visible before a model is trained.
- Measure upstreamFSI uses only pre-OOS characteristic rank persistence.
- Separate the channelsInputs, model sensitivity, and rebalancing create distinct turnover pressure.
- Test out of sampleAcross 15 emerging markets, slower Ridge inputs preserve more net performance at 3× costs.
A real monthly replay needs the trade ledger.
The local research package contains the paper and its measured aggregate results. It does not contain the asset-month export behind those results. This page therefore does not invent a sequence from Sharpe and turnover.
Gross and net Sharpe, turnover, score stability, feature persistence, cost calibrations, model perturbations, and update-rule results.
Market, month, asset ID, score, rank, previous and target weight, trade, spread, return, cost, and gross and net contribution.
- BRA83
- CHN78
- GRC78
- IDN80
- IND81
- KOR86
- MEX93
- MYS90
- PHL87
- POL69
- SAU56
- THA84
- TUR79
- TWN78
- ZAF93
Where does prediction churn enter?
These controls do not build one strategy. Each module opens measured endpoints from a different experiment in the submitted paper.
Experiment ARidge feature sets · 15-market mean · headline OOS windows
What survives the spread?
Accounting and fundamental characteristics move more slowly. They keep similar gross signal while asking the portfolio to trade less.
- pre-OOS rank AC
- 0.97
- score stability
- 0.939
- monthly turnover
- 10.2%
- 12-month IC retained
- 69%
Exact paper endpoints. The switch compares the measured no-cost gross Sharpe with the paper’s 3× capped-spread calibration. It does not interpolate a return path.
Experiment BSeparate sensitivity sample and specification
Keep the slow inputs fixed.
The linear ranking barely changes under the perturbation.
Separate sensitivity exercise. The comparison covers all 15 markets, uses about 102 OOS months per market, and holds the sample and perturbation draws common across models. XGBoost and LightGBM use 100 trees at depth 3; this exercise does not use the top-20 selector. Instability is one minus Spearman rank correlation after adding 0.01-SD noise to every slow characteristic.
Experiment CAll-feature Ridge · measured at 3× costs
Choose what becomes a trade.
Raw monthly ranks preserve the strongest gross signal (0.57), but translate every ranking change directly into the portfolio.
Measured at 3× costs. EMA is selected on the latest 24 fitted months; quarterly is reported as the range across three anchor schedules.
Persistence becomes ranking churn, then cost drag.
The lab above exposes measured states one decision at a time. This view places the 15-market Ridge comparison and pre-OOS FSI validation on one rail.
A stronger gross signal can still lose after trading.
At the same 3× cost calibration, all-feature Ridge starts with the strongest gross Sharpe here, but its turnover reverses the ordering after costs.