X Hedging: towards an eXplainable artificial intelligence option Hedging framework
Authors: Bjørn André Aaslund, Johannes Berge, Ayoub Haida, Ying Ni, Rita Pimentel
Organizations: Department of Industrial Economics and Technology Management, Norwegian University of Science and Technology, Trondheim, Norway. · Department of Business and Mathematics, Mälardalens University, Väster˚as, Sweden.
We propose X Hedging, an alternative to the Deep Hedging framework that addresses computational and explainability limitations of neural-network-based hedging strategies. Deep Hedging leverages neural networks to learn flexible hedging policies but relies on opaque models and typically requires large training datasets, which can complicate interpretation and limit its effectiveness in data-constrained settings. X Hedging takes a step toward addressing these limitations by replacing neural networks with tree ensemble methods, specifically LightGBM, to estimate hedging strategies. This design provides a computationally efficient alternative while preserving strong hedging performance and moving toward more user-centered explainability. To integrate LightGBM into a reinforcement-learning-style hedging framework, we introduce an optimization approach that sequentially trains tree-ensemble models while introducing stochasticity through random training orders. In controlled experiments, X Hedging achieves performance comparable to Deep Hedging while offering greater user-centered explainability, improved computational efficiency, and substantially higher robustness when training data are limited. The numerical results demonstrate that tree-ensemble methods provide an effective computational alternative to neural networks for data-driven dynamic hedging.
Figures & tables
Figure 1: Visual overview of X Hedging and Deep Hedging frameworks. Green boxes show inputs/outputs, blue boxes denote mathematical operations, orange boxes represent subtracted terms in Equation ( 1 ), and yellow boxes indicate the neural network or tree-based model and its output.
S0
K
r
d
σ
T
n
N
1
1
0.0
0.0
0.3
36530
30
200000
Table 1: GBM model parameters for asset price simulation and option parameters.
Mean
Std
JS(DH ∥⋅ )
JS(XH ∥⋅ )
Time (sec)
BS-L
-0.002238
0.005648
0.004041
0.004754
614.0
Call
DH
-0.002007
0.005888
—
0.000146
42616.0
XH
-0.002164
0.005845
—
—
38887.0
BS-L
-0.002667
0.004229
0.002015
0.004007
889.0
Put
DH
-0.003753
0.006070
—
0.000111
45887.0
XH
-0.002260
0.004599
—
—
32346.0
Table 2: Terminal P< summary statistics for ATM options under CTM 2 loss. 1 1 footnotetext: The table reports the mean, standard deviation (Std), and Jensen–Shannon (JS) divergence of terminal P&L distributions under transaction costs.
Figure 2: Histograms of terminal P< distributions for ATM options under transaction costs using the CTM 2 loss function.
Figure 3: Hedging strategies for ATM call option at times k=1,10,15,20,25,29 under transaction costs and the CTM 2 loss function.
Figure 4: Hedging strategies for ATM put option at times k=1,10,15,20,25,29 under transaction costs and the CTM 2 loss function.
Figure 5: Hedging strategies for an ATM put option with data constraints at times k=15,20,25 under proportional transaction costs and the CTM 2 loss function.
Figure 6: One of the X Hedging decision trees visualized at time 9.
Figure 7: Prediction path in one tree for S8 equal to 1.03.
S0
K
r
d
σ
λJ
μJ
σJ
T
n
N
1
1
0.0
0.0
0.3
0.5
−0.1
0.2
36530
30
200000
Table 3: Jump-diffusion model parameters for asset price simulation and option parameters.
Figure 8: Histograms of the P&L T distributions for benchmark (BH) and X Hedging (XH).
Model
Mean
Std Dev
Skewness
Kurtosis
Min
Max
BH
-0.003102
0.018910
-10.830045
173.837940
-0.703028
0.023749
XH
-0.002862
0.019757
-6.600901
90.950098
-0.681716
0.056935
Table 4: Summary statistics of the P< for the benchmark and the X Hedging strategies in a jump-diffusion setting. 1 1 footnotetext: BH denotes the benchmark hedging strategy; XH denotes X Hedging.
Confidence level
VaR
CVaR
(%)
BH
XH
BH
XH
99.5
-0.136173
-0.118569
-0.207796
-0.182217
99.0
-0.083993
-0.075401
-0.157489
-0.138550
97.5
-0.020942
-0.036190
-0.089398
-0.084914
Table 5: VaR and CVaR for the benchmark and the X Hedging assuming different confidence levels. 1 1 footnotetext: BH denotes the benchmark hedging strategy; XH denotes X Hedging.
Appendix figures & tables10 assets
Supplementary material from the paper’s appendix.
Appendix
N. data leaf
N. leaves
N. boost rounds
Learning rate
Early stopping
Iterations
5
10+j
10+5j
0.1
5
15
Appendix
Table 6: Parameters used for the LightGBM models in X Hedging.
Activation function
Optimizer
Initial learning rate
Neurons
Batch size
Epochs
tanh
Adam
0.01
32
1024
10000
Appendix
Table 7: Parameters used for the neural networks in Deep Hedging.
Mean
Std
JS(DH ∥⋅ )
JS(XH ∥⋅ )
Time (sec)
BS
-0.000001
0.005371
0.000273
0.000203
608.0
DH
-0.000004
0.005423
—
0.000277
66303.0
XH
0.000002
0.005505
—
—
45875.0
Appendix
Table 1: Mean, standard deviation and JS-divergence values for P< distributions without transaction costs, using MSE loss function.
Figure 1: Histograms of the P< distributions without transaction costs and using the MSE loss function.
Figure 2: Hedging strategies for ATM call option from times k=1,10,15,20,25,29 without transaction costs and using MSE loss function.
Mean
Std
JS(DH ∥⋅ )
JS(XH ∥⋅ )
Time (sec)
BS-L
-0.002251
0.005659
0.001324
0.001399
309.0
DH
-0.002202
0.005675
—
0.000343
35514.0
XH
-0.002275
0.005763
—
—
21524.0
Appendix
Table 1: Mean, standard deviation and JS-divergence values for P< distributions with transaction costs and MSE loss function.
Figure 1: Histograms of the P< distributions with transaction costs and MSE loss function.
Figure 2: Hedging strategies for ATM call option from times k=1,10,15,20,25,29 with transaction costs and using MSE loss function.
Mean
Std
JS(DH ∥⋅ )
JS(XH ∥⋅ )
Time (sec)
BS-L
-0.001677
0.003397
0.03312
0.032953
565.0
Call
DH
-0.001689
0.003618
—
0.001417
58007.0
XH
-0.001756
0.003664
—
—
30911.0
BS-L
-0.002227
0.005642
0.004922
0.006386
315.0
Put
DH
-0.002130
0.005845
—
0.000344
24532.0
XH
-0.002164
0.005864
—
—
26150.0
Appendix
Table 1: Mean, standard deviation (Std), and JS-divergence values for P< distributions under transaction costs and the CTM 2 loss function for ITM options.
Mean
Std
JS(DH ∥⋅ )
JS(XH ∥⋅ )
Time (sec)
BS-L
-0.001142
0.004263
0.041733
0.029112
622.0
Call
DH
-0.001115
0.004475
—
0.003673
49020.0
XH
-0.001233
0.004558
—
—
35003.0
BS-L
-0.000907
0.003400
0.043407
0.031675
447.0
Put
DH
-0.001043
0.003964
—
0.004696
43102.0
XH
-0.000995
0.003695
—
—
44527.0
Appendix
Table 2: Mean, standard deviation (Std), and JS-divergence values for P< distributions under transaction costs and CTM 2 loss function for OTM options.
Quantitative finance uses mathematical models, statistics, and machine learning to value assets, measure risk, and guide trading and portfolio decisions.
We introduce a deep kernel hedging framework that combines the flexibility of deep learning with the structural inductive bias of kernel methods. The hedging functional is restricted to a reproducing kernel Hilbert space whose kernel is parameterized through a neural network embedding of the input features. The framework minimizes a regularized empirical risk under convex loss functions and can accommodate path-dependent information through truncated time-augmented signature features. We derive a generalized representer theorem for the joint hedging problem, reducing the empirical optimization to a finite-dimensional problem. To further reduce the computational cost associated with large kernel matrices, we develop a scalable random Fourier feature approximation and establish convergence guarantees. The random Fourier parameters are sampled once and remain fixed throughout training, while the deep kernel adapts to market data through the learned neural representation. We evaluate the performance of the proposed deep kernel approach on both synthetic and real data and compare it with standard kernel methods and classical deep hedging architectures. Numerical results indicate competitive and robust hedging performance, particularly in low-data regimes, which highlights the benefits of combining expressive neural representations with the inductive bias of kernel methods.
Jean-Loup Dupret, Donatien Hainaut, Edouard Motte
University of Amsterdam, Amsterdam School of Economics · ETH Zurich, Department of Mathematics, RiskLab · Université Catholique de Louvain, LIDAM–ISBA
We present an end-to-end deep learning framework for systematic options trading that directly embeds hedging behavior through explicit control of portfolio-level risk exposures. While neural networks trained to optimize risk-adjusted performance have been shown to outperform traditional rules-based strategies, such approaches remain agnostic to the sensitivities of the resulting portfolios with respect to specific underlying risk factors. We propose a general training objective that combines a performance-driven loss with a differentiable risk-sensitivity penalty, enforcing neutrality to selected risk dimensions. Unlike reinforcement learning methods that approximate optimal hedging policies via simulated market dynamics, our framework operates entirely on historical data and jointly optimizes risk-adjusted returns and targeted risk constraints in a single learning problem. We instantiate the framework on static delta-neutral straddle portfolios with the penalty directed at first-order directional exposure, and evaluate two penalty variants -- an exposure-normalized penalty and a Greek-ratio drift penalty. Empirical results on Nasdaq 100 equity options demonstrate that appropriately calibrated regularization simultaneously improves out-of-sample risk-adjusted performance relative to an unregularized baseline while reducing realized directional exposure.
Wee Ling Tan, Stephen Roberts, Stefan Zohren
Department of Engineering Science Oxford-Man Institute of Quantitative Finance University of Oxford
Deep hedging learns trading policies from historical or simulated market trajectories, yet under nonstationarity these training paths may not represent future market conditions. We propose WRAP (Wasserstein-Reweighting Adversarial Perturbation), a drift-aware adversarial training framework derived from a two-budget distributionally robust optimization (DRO) formulation. The formulation is anchored to a weighted empirical reference distribution whose fixed baseline weights are chosen to balance sampling uncertainty against temporal drift. Around this reference distribution, the ambiguity set addresses two complementary forms of distributional misspecification by allowing an adversary to reweight the observed trajectories subject to a φ-divergence constraint and perturb their paths subject to an optimal-transport (OT) constraint. We derive a joint first-order expansion in which the leading-order increase over the nominal expected loss decomposes into a reweighting contribution determined by the dispersion of hedging losses across trajectories and a transport contribution determined by the sensitivity of the loss to path perturbations. This expansion yields an explicit finite-dimensional adversarial attack that replaces the distributional inner supremum with a tractable first-order approximation. Across stationary and nonstationary Heston dynamics and a generalized affine diffusion (GAD), the experiments show complementary benefits from reweighting and transport, with joint adversarial training providing the largest gains under nonstationarity.
Philipp J. Schneider, Lukas Looser, Antoine Garin +2
Risk Analytics and Optimization Chair, EPFL, Lausanne, Switzerland · University of Waterloo, Waterloo, Canada