The 41.5% turnover cut is Gelissen and Yuan's most useful trading result, provided turnover means changes in target weights. The 34.13% CAGR deserves more caution: it comes from 22 monthly returns.
The simulator and the trading rule
Gelissen and Yuan fit a statistical model of joint returns to generate trading experience. Each asset has a GARCH-family marginal model with Student-t or skewed-t innovations. BIC selects among them after Ljung-Box screens. A D-vine joins the seven assets through bivariate copulas; all 21 pair copulas are Student-t. Small per-edge neural networks, with two hidden layers of 16 and 8 units, move their correlations through time.
The fitted law supplies 512 joint scenarios each month for a CVaR charge in the reward. It also produces synthetic 24-month paths after a 30-month burn-in, adding to the 61 available historical training paths. A recurrent TD3 (twin-delayed deep deterministic policy gradient) agent uses a 64-unit LSTM to read 30 months of history and choose target weights. The reward takes a CRRA utility increment, with risk aversion 2, and subtracts 0.05 times the 95% CVaR. Net exposure stays at 1, gross exposure cannot exceed 1.5, and positions stay within [-0.20, 0.60]. The trading premise is that simulated experience helps a history-aware policy allocate better than a one-period optimiser.
The assets are three US price indices, three Chinese equity ETFs and a Chinese gold ETF. Daily observations run from December 2013 to July 2026. The last 24 monthly decisions are held out; dropping two short periods leaves 22. The authors also bound policy loss from error in the copula parameters. Under uniform local-derivative and moment assumptions, edgewise error ε enters the value and policy-loss bound at order ε^{1/(2d-1)}. At d = 7, that is ε^{1/13}, with constant 56·A^6. It shrinks too slowly to work as an error bar.
Does 34.13% beat the benchmarks convincingly?
Only partly. The ten-policy ensemble, NN-vine-RL, earns a 34.13% CAGR and a 2.309 Sharpe using a zero risk-free rate, after 10bp turnover costs and financing. The static-vine optimiser earns 31.19% with a 1.704 Sharpe. DCC-GARCH leads the benchmarks on Sharpe at 1.944 and earns 28.69% CAGR. Equal weight returns 24.37% with a 1.702 Sharpe. Its 5.22% drawdown is the smallest in the table, against 5.84% for the ensemble. A period when 1/N compounds at 24% gives a strategy plenty of help.
The authors limit their claim: "these comparisons describe the observed performance of the selected ensemble on this evaluation sample." We did not find a resampling test on the benchmark table itself. Bootstraps address the training-scheme comparisons, with Holm-adjusted p-values from 0.1077 to 0.6406, and the policy-input comparisons, from 0.3198 to 0.6050. The main model has the highest CE in both ablations, each run on the same 22 periods. Beyond a 37-trajectory diagnostic of historical adaptation, we did not find a separate validation window for choosing either the training scheme or the inputs.
A 41.5% cut in target changes
Mixed training reduces monthly turnover from 0.600 under historical-only training to 0.351. CAGR rises from 33.05% to 34.13%. Historical-only training, the paper concedes, has slightly lower volatility and drawdown, plus a higher Sharpe: 2.441 versus 2.309, with drawdown of 4.76% versus 5.84%.
Across the 22 retained periods, mixed training's CE is 33.07% against 32.18%, a 0.89 pp difference whose bootstrap interval is [-8.49, 14.88] pp. Include all 24 periods and the ranking reverses: 30.38% against 30.51%. Mixed training wins on only 4 of 10 seeds, and the median seed difference is -2.81 pp. Growth is a wash. The turnover reduction carries the finding.
Which turnover matters here. Evaluation uses the change in target weights, ||w_t - w_{t-1}||. Fixed targets therefore score zero even when price drift requires trades to restore them; the paper notes this for equal weight's 0.000 entry. Training instead charges drift-aware turnover against pre-trade weights. The pre-training diagnostics use that measure too, with mean turnover falling from 0.933 to 0.627. Thus the evaluation tables' 41.5% reduction measures target changes and understates trades needed to undo drift for a policy with stable targets. Under that accounting, transaction and financing costs together reduced the ensemble's total return by 1.61 pp, leaving a modest dollar gap. Drift-aware accounting could change the ratio between training schemes. A rerun on that basis would show how much of the headline cut remains.
Averaging changes the book
Averaging the ten policies' targets brings average gross exposure down from 1.399 to 1.059. The paper says this cancels 85.31% of gross exposure above one and cuts shorts by the same proportion. Individual agents hold a levered long-short book; the implemented portfolio is close to long-only. Financing charges fall with that shift.
The input ablation helps locate the vine's contribution. Give the policy vine correlations, Kendall's τ and tail dependence, or scalar CVaR, and CE falls by between 1.84 and 8.02 pp relative to the compressed state that omits them. None of the three differences survives Holm adjustment (p 0.3198 to 0.6050). On this sample, the vine helps through simulated paths and the risk charge rather than through those policy inputs.
Settings left open
The configuration table specifies much of the learning loop: actor learning rates of 3×10^-5 in pre-training and 2×10^-5 in fine-tuning, target noise of 0.10 clipped at 0.25, and 512 scenarios whose 26 largest losses are averaged. We could not find values for several settings that affect behaviour. They include short and cash borrowing rates in the cost multiplier, entropy and leverage weights in the actor objective and their reference coefficient, softmax temperature, the bound on each edge's correlation departure, and the moment vector and weights used in correlation recalibration. The authors say the simulation error and optimisation gap in their policy-loss bound remain unquantified. These settings require judgment; the leverage weight shapes gross exposure before averaging starts.
Why we did not rerun it
We do not have price series for the four Chinese-listed ETFs (sh510050, sh510880, sz159915, sz159934). Substituting another market would change the fitted vine. Correlations within the Chinese block range from 0.621 to 0.857, versus 0.198 to 0.321 with US indices; equity-gold correlations range from -0.093 to 0.052. Different assets would produce a different 21-edge dependence model and a different strategy.