LiMT improves CSI300 IC by about 7% over XGBoost. The cleaner result is the ranker. The headline portfolio return, 10.01% versus 3.99% under equal weighting, depends on a sizing parameter swept across the test years.
The ranker, then the sizing rule
Yang, Lin, Qi and their co-authors train a daily cross-sectional ranker for Chinese A-shares. For every CSI300 or CSI500 constituent, it receives 21 days of Qlib's Alpha158 price-volume features.
The first block is the Market Regime Encoder. At each of the 21 historical dates, it applies attention across stocks. Attention then runs through each stock's own history, with only the final step retained. The authors' argument is that day-specific co-movement among peers should be observed before averaging can wash it out.
Liquidity-Driven Learning, the second block, uses a multi-gate mixture of experts. It predicts three cross-sectionally z-scored targets for the session two days ahead: close-to-close return, volume shock and the high-minus-low range over VWAP. Volume shock is log volume less its previous five-day mean. Two per-stock sigmoid gates pass the volume and volatility features into the return head. Return rank supplies the trading signal. The auxiliary targets are intended to improve that ranking and guide position size.
Adaptive Portfolio Optimization, or APO, handles the sizing. It buys the top return decile and shorts the bottom decile. Within each side, APO combines two min-max weights through a parameter r. One uses predicted traded value, while the other uses negative predicted volatility. Every position is capped at 2% of daily turnover. The training period runs from 2008 to 2014, validation covers 2015 to 2016, and testing spans 2017-01-01 to 2020-08-01, using three seeds.
A 7% CSI300 IC advantage
On CSI300, LiMT records test IC of 0.0575. XGBoost, the strongest baseline, reaches 0.0538. LiMT's ICIR is 0.4462, compared with MASTER's 0.4321. The CSI500 figures make a stronger case: IC is 0.0496 against MASTER's 0.0435, while ICIR is 0.452 against 0.3962 for GRU. Removing cross-stock attention reduces CSI300 IC to 0.0493, supporting the encoder design.
The evidence for the multi-task section carries less force. Without the auxiliary targets, IC declines to 0.0543 and ICIR to 0.4035. The paper reports the deterioration from this ablation as 1.39% to 7.80%, although the ICIR reduction is 9.6%.
Another CSI300 ablation removes the entire LDL block, including its experts and gated transfer, and sends the encoder output directly to the three heads. IC remains at 0.0568. Excess return rises to 0.1845, exceeding the full model's 0.1809. The stripped version remains behind on ICIR, at 0.4342 versus 0.4462, and on IR excess, at 1.9993 versus 2.0414. Its advantage is confined to AR excess. Even so, the paper describes gated transfer as "a key part of the improvement."
The auxiliary heads are strongest on the easier forecasts. Volatility IC reaches 0.5157, and volume-shock IC is 0.3435. Range volatility persists, so a 0.5157 IC there measures a range-volatility forecast. The return-side contribution is the change from 0.0543 without the auxiliary targets to 0.0575 with them.
The excess-return columns warrant caution. LiMT's 0.1809 annualized excess return on CSI300 comes from Qlib's top-50 long portfolio, with 20 daily replacements. We did not find a cost assumption for that simulation. A 7% IC lead over XGBoost accompanies more than twice the excess return, 0.1809 against 0.0820. DoubleEnsemble is the strongest excess-return baseline, at 0.1158 versus LiMT's 0.1809. Turnover or costs could explain such a large separation between ranking quality and P&L, and Table I does not distinguish them. We also did not find a statement about point-in-time index membership.
How APO reaches 10.01%
After 10 bp per trade, APO raises CSI300 long-short annualized return from 3.99% to 10.01%. Sharpe climbs from 1.22 to 1.86 relative to equal weighting.
The authors acknowledge the price. Drawdown increases from 3.73% to 4.94%, and turnover rises from 20.73% to 24.98%. They characterize this as a deliberate trade-off "rather than mechanically lower-risk allocations." The description is fair: Sortino improves from 1.29 to 2.00, while Calmar moves from 1.07 to 2.03. The remaining issue is the source of the gain.
The formula gives a fairly direct answer. Predicted traded value is reconstructed by adding each stock's trailing five-day mean log volume back to the forecast, then multiplying by VWAP. Its level therefore reflects trailing dollar volume, with the predicted shock adjusting it. At r=0.1, this component receives 90% of the weight. Under our reading, APO's best setting comes close to weighting each decile by recent dollar volume, introducing a size and liquidity tilt.
The three competing rules use equal weight, inverse volatility and rank-based sizing. None uses trailing dollar volume. That missing comparison would separate the volume forecast's information from the tilt itself. The abstract presents the backtest as evidence "that the multi-task forecasts translate into deployable portfolio gains." A dollar-volume baseline would test precisely that claim.
The choice of r raises a separate concern. The sensitivity table varies r from 0.1 to 0.9 across the same 2017 to 2020 period used for the headline result. Returns then decline monotonically, from 10.01% to 2.62%. The selected setting occupies the grid's edge and is also its best row. The paper describes no validation-period selection of r, so the choice reads as coming from the headline window itself. Values below 0.1 were not tested. Table IV is described as evidence that "a mild liquidity preference is most effective", although r=0.1 assigns 0.9 to the liquidity term in the formula.
A regression provides further support in the paper. The series called "the underlying model-induced long-short spread" earns daily Fama-French five-factor alpha of 0.35% with p<0.01. Across 252 days, that is about 88% a year before compounding, compared with 10.01% for the portfolio. We could not identify the A-share factor definitions or determine whether the spread is net of costs. And the regression uses a series other than the APO portfolio, limiting what it says about improved execution rather than greater style exposure.
Shorts, participation and fills
Each side is allocated CNY 1 billion. The frictions we found are 10 bp per trade and a 2% participation cap. Because the cap uses day-t turnover, the relevant figure is available before trading. We found no assumption for borrow cost or short availability. In A-shares, implicitly assuming that a billion yuan of bottom-decile CSI300 positions can be shorted is consequential.
OSR, named order success rate in the paper, increases from 32.13% to 38.00%. We found no formula for the measure. If OSR represents fill rate, most orders fail under every tested rule.
A US adaptation
We cannot rerun the paper. CSI300 and CSI500 point-in-time membership and A-share OHLCV are unavailable to us. Our US price history also starts around 2010, later than the paper's 2008 starting point.
We are building a US version using a liquid universe chosen by point-in-time market cap, with Alpha158-style features reconstructed solely from OHLCV. The resulting figures will apply to that adaptation. None of the paper's Chinese results carries over to it.
The per-date cross-stock encoder and the three-target head are the portable parts. Our main test is whether APO at r=0.1 can beat straightforward trailing dollar-volume weighting within the same deciles after US costs and borrow are charged.
A win would show that predicted volume shock contains sizing information in our US universe. A match would leave our US run without evidence that the volume forecast produced the sizing gain, while saying nothing conclusive about what generated 10.01% on CSI300.