Yesterday's close beat everything Zhong, Wang, Wang and Feng trained on 29,514 out-of-sample bond-days of Chinese convertibles. The small net's MSE improvement is capped at 0.45%, leaving persistence standing. The paper goes further when it attributes pooling's advantage to shrinkage and the losses to contract clauses. Neither explanation has the same support as the forecasting result.
A clean test of tomorrow's close
Chinese exchange-traded convertibles come with issuer calls, investor puts and downward conversion-price resets. Their path dependence has attracted considerable machine-learning work. Much of that work fits model values to contemporaneous prices or ranks weekly cross-sectional returns. The weekly-return study by Li and coauthors, cited in this paper, is an example. Here the question is more specific: can a network predict tomorrow's closing price better than today's close?
The sample contains Wind daily closes for 100 A-share convertibles from April 2022 to December 2025. The authors train on 2022-2023 and validate on 2024H1. Testing covers July 2024 to December 2025, with 90 bonds and 367 target dates. Each model forecasts the one-day price change, then adds that forecast to the current close. The two MLPs are a small network with widths [64,32] and a large four-layer network. Each is trained on five seeds, whose forecasts are averaged, and neither receives a bond identifier. Historical drift, a pooled AR(1), Ridge and Elastic Net supply the baselines.
The care taken with the test is the paper's strongest contribution. The authors purge rows whose information and target dates fall on opposite sides of a split (98 plus 90 rows). They refit scalers within each training fold and lag stock and Wind valuation fields by an extra day. Errors are clustered by target date because bonds trading on the same date share shocks. Inference uses a Newey-West test alongside a 10-day block bootstrap of out-of-sample R² relative to the random walk.
Does either network beat persistence?
Neither does. Against random-walk RMSE of 2.3577 CNY, the small ensemble records 2.3619 (R²_OOS -0.0035, p=0.453). The large ensemble records 2.3645 (-0.0057, p=0.248). At the one-sided 95% level, the upper bounds on R²_OOS are 0.0045 and 0.0022. Those bounds rule out MSE improvements greater than about 0.45% and 0.22% relative to the random walk. The authors qualify this uncertainty: it is conditional on the fitted forecasts and leaves out model-selection and retraining noise. The true band is therefore wider, by an amount they do not estimate.
The simpler alternatives fare worse, too. Drift (p<0.001) and AR(1) (p=0.018) both lose significantly to persistence.
Directional accuracy offers no rescue. Up days account for 52.07% of the test, while the price-lag ensembles get direction right on 50.53% and 49.57%. They predict declines on 63.69% and 73.45% of bond-days. All of these are forecast statistics. The paper tests no trading rule and charges no costs.
Pooling's small, unresolved advantage
The abstract describes a "variance-stabilizing shrinkage effect," citing lower test RMSE on 79 of 90 bonds than with bond-by-bond networks. That count of 79 uses mean seed-level RMSE for each bond. Compare the two five-seed ensembles directly on identical rows and the RMSEs are 2.3619 versus 2.3655. Pooling leads by 0.153%, with p=0.324 and an interval that crosses zero.
Training effort also differs under the default recipe. Pooled fits receive 980 to 1,323 optimizer updates; single-bond fits receive 16 to 100. Equalizing updates per fit leaves pooling with edges of -0.018%, 0.190% and 0.276% at 50, 100 and 250 updates. Every interval includes zero.
The body of the paper is clear about the limit: "These checks support a forecast-error comparison but do not identify changes in parameter variance." The authors also say pooling "has a favorable point estimate" and its advantage "is not statistically resolved." Their body text gets the distinction right. The mechanism asserted in the abstract and conclusion goes beyond it. A median per-bond improvement of 0.0102 CNY, with pooling ahead on 79 of 90 bonds, is plausible. Its comparator remains a set of single-bond fits that beat the random walk on just 12 of 90 bonds.
Five bonds account for 74.4% of positive excess error
In the abstract and conclusion, five bonds are said to generate 74.4% of the model's total excess squared error. Table 8 uses a narrower denominator: positive excess error only, with those bonds contributing 848.99 of 1,141.3 CNY². Across all 90 bonds, net excess error is 581.20 CNY². The five therefore lose more than the panel's entire net 581.20. Bonds on which the model outperformed persistence saved a total of 560.20 CNY².
Absolute CNY is the unit of loss. With one-day jumps of 25.30 to 50.13 CNY, these bonds receive heavy weight in squared yuan errors by construction. We did not find the decomposition repeated in return-scaled terms. Random-walk RMSE rises from 2.0060 in 2025H1 to 3.0068 in 2025H2, consistent with the same account of a few volatile names driving the result.
The causal explanation has less support still. The paper says jumps "typically coincide" with equity speculative frenzies, reset announcements or call deadlines. We did not find a table matching event dates to the jump days. The paper does not separate equity-driven jumps from clause-driven jumps. A regressor fed momentum would overshoot on either. This panel suggests clause-driven failure as a hypothesis; the paper presents it as a finding.
Persistence stays on the desk
Two results should alter how desks assess these models. Start with the walk-forward results. In 2024H2, R²_OOS is +0.0045 and +0.0048 (p=0.414, 0.324). Both turn negative afterward, reaching -0.0078 and -0.0238 in 2025H2. A decision based on that good half-year would have shipped a model that subsequently lost.
The other result concerns the prediction target. Fitting price levels rather than changes takes the same networks from about -0.004 to between -0.1042 and -0.5212. A quoted price-level R² for a convertible model measures how closely the network copies the current close.
For ten thin-history bonds, the leave-one-bond-out test produces R²_OOS of 0.0130, with p=0.090 and an interval of [-0.0008, 0.0248]. The authors describe this as exploratory. The description fits.
The scope remains limited. These tests cover MLPs and linear models fed lagged prices, technical indicators and lagged valuation and equity fields. As the authors acknowledge, trees, transformers and same-day equity moves remain untested. So does any strategy after costs.
We could not rerun the study ourselves. We hold no Chinese convertible bond price series, and substituting US equities would remove the individual-bond jump dynamics at issue here.
A model using dated call and reset countdowns that beats persistence on those five bonds specifically would change our view.