Set the risk penalty to zero and CAST collapses. NASDAQ Sharpe falls from 0.523 to -0.884. Maximum drawdown jumps from 11.4% to 80.6%, leaving the $1000 account at $200. CSI300 tells the same story: 81.6% drawdown and $240 left.
The authors acknowledge as much. Section V.A identifies the risk penalty as the most significant of the controller's three parts, while the abstract credits forecast uncertainty with controlling drawdown. Yet the title and contribution list continue to sell the cross-asset filter.
Peng, Khushi and Poon call their two-stage daily system CAST. Stage one forecasts each stock with three parallel Kalman filters, using integrated random walks of orders 1, 2 and 3. Credibility weights blend their outputs. Those weights reflect each filter's accumulated squared error on the forecast issued L days earlier. Gradient training never enters the system. Parameters are calibrated once on 2005-2010, then frozen.
CoKF extends the model across assets by stacking all 30 into a joint state. Coupling enters only through the process-noise covariance. Each off-diagonal entry equals rho_ij times the two innovation scales, with rho defined as the Pearson correlation of daily price changes estimated on pre-2010 data. Observation noise remains diagonal. The construction is defensible: shared market shocks enter the latent value process, while measurement error stays with each print.
Stage two determines whether the system makes money. At every close, a per-asset model predictive controller plans a short trade sequence. Its objective adds u times the predicted price change across the horizon, then subtracts lambda times the sum of |u| times omega_l. The l-step forecast standard deviation, omega_l, comes from the credibility-averaged transition matrix.
Two constraints govern the plan. The trades must sum to minus the latest trade, so the position unwinds by the horizon's end. No individual trade may exceed 50% of account value. The problem linearizes to an LP. Only the first trade executes, and the controller solves again at the next close.
Thirty separate accounts, one for each stock, with no netting or portfolio-level allocation.
This amounts to volatility targeting under a different label. The volatility estimate comes from the filter's forecast transition instead of realized returns.
The controller carries the strategy
Single-order filters fare badly in the predictor ablation. On NASDAQ, KF r=2 alone records Sharpe -0.555 with 64.3% drawdown. The r=3 version reaches -0.416 and 53.8%. The r=1 random walk generates no directional signal, so it never trades. The authors describe that outcome as intended behaviour rather than failure.
Blending the three changes the result. Their single-asset CKF produces NASDAQ Sharpe 0.339 at 10.3% drawdown. Adaptive order selection is contributing something real.
The paper concludes that every MPC component is necessary, and its table supports that view. Reducing the controller to a single-step decision sends Sharpe from 0.523 to -0.271. Drawdown rises from 11.4% to 46.7%, with $627 remaining. Removing MPC entirely produces -0.442 at 57.0% drawdown and leaves $453.
Both losses are large. The lambda term alone accounts for 1.4 of Sharpe, far more than either change, even though the headline architecture does not contain that term. We did not find a comparison against fixed-fraction sizing or realized-volatility targeting anywhere in the experiments section. The ablation therefore shows that uncertainty-scaled sizing beats no sizing. Nobody doubted that.
Coupling helps drawdown, unevenly
Across the four 30-stock baskets over 2010.01-2025.04, CAST improves on CKF drawdown in three. CSI300 comes in at 15.8% against 21.8%, TPX100 at 13.4% against 15.0%, and Global30 at 3.1% against 10.7%. NASDAQ is the exception, with 11.4% against 10.3%.
Final value rises in two of the four markets. NASDAQ finishes at $1633 against $1236, while CSI300 ends at $1272 against $1117. TPX100 falls to $1152 against $1310, and Global30 to $1133 against $1178. Sharpe improves in three of four, including NASDAQ, where it moves from 0.339 to 0.523. During the 2022 rate-hike window on NASDAQ, CKF limits the loss to 2.6%; CAST gives up 10.3%.
Correlation coupling is chiefly a drawdown trade. It sacrifices return in half the markets tested. The result is defensible, though it differs from the return story implied by the abstract.
A consistency problem remains. The paper begins with a figure showing that stock-correlation distributions shift between periods, then estimates rho once on pre-2010 data and holds it fixed for fifteen years. The predictor adapts online. The coupling matrix stays frozen.
Returns few traders will quote
Over 15.25 years, CAST grows $1000 to $1633 on NASDAQ, $1272 on CSI300, $1152 on TPX100 and $1133 on Global30. That works out to 0.8% to 3.2% a year. A NASDAQ Sharpe of 0.523 is a real result, and the drawdowns are genuinely small. Global30 records 3.1% over the full window, the lowest among all 17 methods tested.
Exposure, however, is sparse. The crisis table makes this plain. During the 2020 COVID crash, CAST loses 0.9% on TPX100 and 1.0% on Global30. Buy-and-hold loses 25.0% and 25.8%. Losing one percent while the index loses twenty-five mostly reflects an off-switch.
The authors phrase their claim carefully. CAST lies on the Pareto frontier for final value against drawdown, so no baseline beats it on both measures at once. Their table supports that statement. RSR reaches $2196 on NASDAQ with 12.7% drawdown. Informer makes $6088 on Global30 at 36.2%, compared with CAST's $1133 at 3.1%. The authors also fairly observe that many low-drawdown baselines achieved those losses by giving up enough return to enter negative Sharpe territory.
Readers never see CAST's full-window return beside buy-and-hold. The market benchmark appears only within the two crisis windows.
The authors describe the backtest as frictionless: "zero transaction costs and slippage, no leverage." The controller trades at every close, and each trade may reach 50% of account value. Its planned sequence unwinds the latest trade by the horizon's end. The paper reports no turnover figure. With annualized returns of 0.8% to 3.2%, the cost budget is roughly zero.
Our test
We are trading a US-equity version of the same mechanism on daily bars under realistic execution assumptions. The CSI300, TPX100 and Global30 figures above do not carry over to a US-only book, so this is a separate experiment rather than a replication of their work. The mechanism transfers readily because it requires only a panel of daily closes and a correlation matrix.
Turnover relative to the penalty will decide the result. Their 0.523 was measured on NASDAQ with zero transaction costs and zero slippage. A costed book cannot be held to that bar. Suppose uncertainty-scaled sizing survives realistic execution and keeps drawdowns in the low teens. The MPC layer would then merit a place above whichever forecaster you already trust.
If it does not survive, CAST joins the long line of frictionless backtests we have flagged before. We described the same failure when we reviewed seventy-seven agentic trading papers with one transaction-cost model between them.
Whether CoKF is the right predictor remains open under the paper's own ablations.
One thing would change my mind about the cross-asset term. Re-estimate rho on a rolling window. If that version beats fixed-rho CAST on drawdown and closes the NASDAQ gap, where coupling costs 11.4% against CKF's 10.3%, the coupling has earned its place.