CAST

Abstract

Managing drawdown, the peak-to-trough decline in an investment portfolio’s value, is a precondition for long-term survival in practical investment management. However, mainstream stock forecasting methods predominantly optimize returns or Sharpe ratios under the independent and identically distributed (i.i.d.) assumption. Real markets do not follow this assumption, triggering catastrophic drawdowns.

We propose a cross-asset state-space trading system (CAST), consisting of two components: The predictor, Cross-Asset Collaborative Kalman Filter (CoKF), estimates each asset’s latent state online, coupling all assets through their correlations and adaptively fusing multiple integrated-random-walk orders. The controller, Model Predictive Control (MPC), converts the predictor’s forecast into trades, using forecast uncertainty as an explicit risk penalty that controls drawdown.

We evaluate CAST on four real-world stock markets over a 15-year test window and show that it consistently occupies the return–drawdown Pareto frontier, achieving strong risk-adjusted performance while maintaining substantially lower maximum drawdown than competitive baselines. A stress test across crisis periods further demonstrates robust behavior under market shocks and distribution shift. Because the predictor and controller interact only through the predicted price path, both are plug-and-play, making CAST a modular, interpretable trading system.

Poster

Datasets

4real-world stock markets

30stocks per panel, daily close

20yrsof data, Jan 2005 – Apr 2025

15yrsout-of-sample test window

United States NASDAQ NASDAQ Composite constituents

China CSI 300 Large-cap A-shares

Japan TPX100 Tokyo Stock Exchange blue-chips

Cross-currency Global30 Five currency zones: USD, EUR, JPY, GBP, CNY

2005–2010 · calibrate

2010.01 – 2025.04 · out-of-sample test, never refitted

The 2008 crisis sits in the calibration segment; the test window brings two different ones — the 2020 COVID-19 liquidity shock and the 2022 rate-hike repricing. Parameters are calibrated once on pre-2010 data and never refitted. Each panel is one parquet file of prices plus a JSON ticker list; Global30 prices are normalised to their first-day value before backtesting, the other three panels use raw prices.

Method

CAST architecture: the CoKF predictor produces an L-step forecast and its uncertainty; the MPC controller trades only when the forecast outweighs the uncertainty — **Overall framework.** The predictor (CoKF) emits an L-step price forecast P̄_k+l|k together with its forecast uncertainty ω_l; the controller (MPC) plans the whole trade sequence, penalises every trade by λ·|u|·ω_l, and commits only u_k+1 before re-planning at the next close. The two modules share nothing but this forecast–uncertainty pair, so either is plug-and-play.

PredictorCoKF

Each close price is treated as a noisy reading of a hidden intrinsic value, tracked by a Kalman filter. It updates recursively from the latest price — no retraining, constant work per step — and its error covariance reports how uncertain it is.

Multi-order. Integrated-random-walk filters of order r = 1, 2, 3 (constant, linear and parabolic local paths) run in parallel. Each is scored by its recent L-step error, and a closed-form credibility weight μ shifts trust between orders every day, so the model adapts to regime shifts online.

Cross-asset. All assets share one joint filter coupled through the process-noise covariance Q̃, whose off-diagonals are ρ_ijσ_v⁽ⁱ⁾σ_v^(j): shared market shocks move the latent dynamics, not idiosyncratic measurements. With ρ = I it reduces exactly to independent per-asset filters.

ControllerMPC

One risk-isolated controller per asset. On a receding horizon of L = 7 trading days it plans the trade sequence u_k+1, …, u_k+L−1 maximising predicted horizon profit net of the uncertainty penalty, under two constraints: the planned trades must sum to −u_k (unwinding today’s trade by the horizon end), and no single trade may exceed half the account value (β = 0.5). Only the first trade is executed; the plan is rebuilt at the next close with fresh prices.

Uncertainty as risk. The penalty weight ω_l is the l-step forecast standard deviation, built from the same credibility weights that produced the forecast. When uncertainty rises, the penalty grows and positions shrink — the system de-risks automatically as a crisis sets in.

With slack variables for |u| the problem is a linear program, solved in sub-millisecond time.

Results

Table II: comparison of CKF and CAST with 15 baselines on NASDAQ, CSI300, TPX100 and Global30 — **Main results (Table II).** CKF and CAST against 15 baselines — Transformer, Linear/MLP, financial, distribution-shift-adaptation and SOTA stock forecasters — on all four markets. FV: final value from $1000; MDD: maximum drawdown (lower is better); CR, ASR, SoR: Calmar, annualized Sharpe and Sortino ratios (higher is better).

Supplementary · Table III

Maximum drawdown during two crises

	COVID-19 crash (2020)				Rate-hike crisis (2022)
Method	NASDAQ	CSI300	TPX100	Global30	NASDAQ	CSI300	TPX100	Global30
Market	0.340	0.159	0.250	0.258	0.252	0.172	0.150	0.336
BLSW	0.265	0.043	0.045	0.148	0.061	0.071	0.039	0.066
CSM	0.127	0.042	0.015	0.031	0.186	0.065	0.057	0.166
AlphaStock	0.052	0.032	0.032	0.033	0.070	0.059	0.036	0.064
MetaTrader	0.162	0.040	0.041	0.048	0.096	0.057	0.067	0.076
FreQuant	0.186	0.045	0.030	0.094	0.116	0.063	0.067	0.090
CKF	0.045	0.057	0.055	0.009	0.026	0.086	0.033	0.018
CAST	0.071	0.035	0.009	0.010	0.103	0.077	0.016	0.017

Crisis stress test. Bold red marks the lowest (best) MDD and underline the second lowest. Buy-and-hold loses 15–34% of capital across these crises; CKF or CAST posts the lowest drawdown in 6 of the 8 columns. Baselines: traditional strategies (Market, BLSW, CSM) and RL-based portfolio methods (AlphaStock, MetaTrader, FreQuant).

Citation

@misc{peng2026cast,
  title        = {{CAST}: A Cross-Asset State-Space Trading System
                  for Drawdown Control in Stock Markets},
  author       = {Peng, Yu and Khushi, Matloob and Poon, Josiah},
  year         = {2026},
  eprint       = {2609.14205},
  archivePrefix= {arXiv},
  primaryClass = {cs.CE},
  doi          = {10.48550/arXiv.2609.14205},
  url          = {https://arxiv.org/abs/2609.14205},
  note         = {Accepted at IEEE International Conference on Data Mining (ICDM 2026)}
}

CASTA Cross-Asset State-Space Trading Systemfor Drawdown Control in Stock Markets