Source-linked AI summary
Deep Hedging
Hans Bühler, Lukas Gonon, Josef Teichmann, Ben Wood
TL;DR
The paper addresses hedging and pricing in incomplete markets with transaction costs, impact, liquidity constraints, and risk limits. It uses neural-network strategies with convex risk measures, establishes approximation properties, and evaluates the approach through out-of-sample experiments in a Heston model.
Problem
Incomplete markets with frictions require minimal-price and hedging criteria that account for costs, constraints, and nonlinear dependence on the overall book.
Method
The paper approximates constrained hedging strategies with neural networks using market information features and convex risk measures.
Results
The analysis uses out-of-sample performance to compare neural-network hedging with a Heston-model benchmark, assess transaction-cost effects, and test scalability to higher dimensions.
Takeaways & Limitations
Neural-network strategies are theoretically well-founded through universal approximation and can be calculated efficiently for hedging under convex risk measures.
Takeaways & Limitations
Modeling the effect of trading in one hedging instrument on other instruments remains a real challenge, especially for options.
Abstract
from arXiv · showhide
We present a framework for hedging a portfolio of derivatives in the presence of market frictions such as transaction costs, market impact, liquidity constraints or risk limits using modern deep reinforcement machine learning methods. We discuss how standard reinforcement learning methods can be applied to non-linear reward structures, i.e. in our case convex risk measures. As a general contribution to the use of deep learning for stochastic processes, we also show that the set of constrained trading strategies used by our algorithm is large enough to $ε$-approximate any optimal solution. Our algorithm can be implemented efficiently even in high-dimensional situations using modern machine learning tools. Its structure does not depend on specific market dynamics, and generalizes across hedging instruments including the use of liquid derivatives. Its computational performance is largely invariant in the size of the portfolio as it depends mainly on the number of hedging instruments available. We illustrate our approach by showing the effect on hedging under transaction costs in a synthetic market driven by the Heston model, where we outperform the standard "complete market" solution.
1. Introduction
The paper addresses derivative hedging when real-market frictions make complete-market methods inadequate, using model-free neural-network strategies optimized for convex risk measures. It develops and evaluates deep hedging in Heston experiments, including transaction costs, scalability, and comparisons with established approaches.
- Motivation: Transaction costs, market impact, liquidity constraints, and risk limits make pricing and risk dependent on the overall derivatives book rather than linear complete-market formulas.A new trade may reduce risk in one direction and receive a more favorable price, an effect described as having an “axe”.
- Motivation: Complete-market models remain prevalent because efficient alternatives rarely scale across large portfolios or remain independent of market dynamics.
- Approach: Deep hedging models trading decisions as neural networks whose features can include instrument prices, trading signals, news analytics, and past hedging decisions.
- Approach: The approach is model-free and can incorporate transaction costs, liquidity constraints, bid/ask spreads, and market impact that depend on scenario features.
- Experiments: The framework applies reinforcement learning to derivative hedging under convex risk measures and is evaluated using out-of-sample performance in Heston-based experiments.Experiments examine benchmark comparison without transaction costs, proportional transaction costs, and scalability to higher dimensions.
- Related work: The article’s novelty is covering derivatives, especially over-the-counter derivatives without observable market prices, rather than only classic portfolio optimization or algorithmic trading.
2. Setting: Discrete time-market with Frictions
The framework represents a finite-horizon market with derivative liabilities, multiple hedging instruments, information-driven decisions, trading constraints, and explicit friction costs. It supports proportional, fixed, cross-asset, and market-impact cost models, while recognizing cross-instrument impact as a practical modeling challenge.
- Information: Market information I_k includes costs, prices, implied volatilities, news, balance-sheet information, trading signals, and risk limits, generating the filtration used for decisions.The history I_0,...,I_k forms the richest available feature set for a decision at t_k.
- Market and claims: The market is a finite discrete-time model with d hedging instruments, adapted mid-prices, and no requirement for an equivalent martingale measure.Hedging instruments may include liquid secondary assets such as options, not only primary assets such as equities.
- Market and claims: The liability Z is an FT-measurable portfolio combining liquid and OTC derivatives, and no classic derivative pricing model is needed to value it or compute Greeks.
- Strategies and constraints: Trading strategies are adapted holdings processes subject to restrictions arising from liquidity, asset availability, and trading rules, including prohibitions before an option becomes tradable.
- Friction models: Trading costs are accumulated in terminal portfolio value, with the setup allowing proportional costs, fixed costs, and cross-surface costs based on option Delta and Vega.The framework also permits true market impact, where trading decisions affect asset distributions.
- Friction models: The main modeling challenge for market impact is representing how trading one hedging instrument affects other instruments, particularly options.
3. Pricing and hedging using convex risk measures
In incomplete markets with frictions, convex risk measures define acceptable positions and prices through optimized hedging. The framework also recovers familiar indifference-pricing results in important special cases.
- Pricing and hedging under frictions: Complete-market replication is generally unavailable with trading restrictions and transaction costs, so an optimality criterion is needed to define acceptable minimal prices.The minimal price is the cash added to implement the optimal hedge while making the overall position acceptable under costs and constraints.
- Convex risk measures: A convex risk measure is monotone, convex, and cash-invariant, respectively encoding preference for more favorable positions, diversification, and cash offsets.Cash invariance implies that ρ(X) is the least cash amount needed to make position X acceptable.
- Convex risk measures: The hedging functional π is monotone decreasing and cash-invariant, and becomes a convex risk measure when transaction costs and the admissible strategy set are convex.The paper defines optimal hedging as minimizing π over admissible strategies.
- Indifference pricing: The indifference price p(Z) is the cash charge making the liability position −Z equivalent in risk to doing nothing, through π(−Z + p(Z)) = π(0).Without trading restrictions and transaction costs, this price coincides with a replicating-portfolio price when replication exists.
- Indifference pricing: For an attainable claim in a frictionless unrestricted market, the indifference price equals its replication price p0.The result assumes CT ≡ 0 and H = Hu.
- Extensions: The framework can also optimize expected terminal loss when a price p0 is specified exogenously, including optimal hedging under a capital constraint.The objective is inf over admissible strategies of E[ℓ(−Z + p0 + (δ · S)T − CT(δ))].
- Arbitrage and relevance: Arbitrage can make π(X) equal −∞, but classic arbitrage need not make a market irrelevant when the risk measure accounts for a non-zero probability of no gain.Under the stated extreme risk measure, the market remains relevant with π(0) = 0.
- Special cases: Choosing the entropic risk measure makes the framework's indifference price equal to exponential-utility indifference pricing.The paper also connects loss-function constructions to convex risk measures and optimized certainty equivalents.
4. Approximating hedging strategies by deep neural networks
The paper approximates constrained hedging strategies with parameterized neural networks, reducing an infinite-dimensional optimization to finite-dimensional parameter optimization. Universal approximation supports arbitrarily accurate strategy approximation, while the framework extends to convex risk measures and practical trading constraints.
- Motivation: Neural networks are the paper’s central parametric family for approximating hedging strategies and enabling efficient numerical optimization.The approach is presented as theoretically grounded and computationally efficient.
- Neural-network construction: A feed-forward network composes affine maps with componentwise activation functions across input, hidden, and output layers.Weights and biases parameterize the affine maps, and nonzero weights characterize network complexity.
- Approximation theory: Bounded, non-constant activation functions give neural networks dense approximation in Lp for finite measures, with continuous activations also dense under uniform convergence on compact sets.These universal approximation properties motivate using increasingly rich network classes.
- Hedging-strategy parametrization: The semi-recurrent strategy class uses current and past market information together with the previous trading position as network inputs.Strategies are represented as δk = Fk(I0, . . . , Ik, δk−1), with Fk drawn from a finite-dimensional network class.
- Optimization formulation: Replacing the unconstrained strategy class with HM yields the finite-dimensional objective πM(X), optimized over neural-network parameters.The constrained problem is represented through the transformation H ◦ δ, and the resulting optimization is finite-dimensional.
- Theoretical guarantee: Universal approximation implies that strategies in H can be approximated arbitrarily well by strategies in HM, and the neural-network price converges to the exact price.The result is stated for the network price πM(−Z) − πM(0) converging to p(Z).
5. Numerical experiments and results
The experiments evaluate deep hedging against benchmarks, under different convex risk preferences, proportional transaction costs, and higher-dimensional settings. In Heston-based tests, neural networks closely approximate optimal or model hedges, while the method also verifies transaction-cost asymptotics and shows potential for high-dimensional hedging.
- Experimental design: The experiments assess benchmark accuracy, transaction-cost effects, and scalability using out-of-sample performance.The study varies risk preferences, proportional transaction costs, and the number of hedging instruments.
- Experimental design: The Heston setup uses a 30-day horizon with daily rebalancing, trading in the stock and an idealized variance swap.The market is driven by a stochastic volatility model with two liquidly tradable hedging instruments.
- Risk preferences: Higher risk aversion produces a more flattened call-spread hedge, corresponding practically to a barrier shift.The comparison considers α = 0.95 and α = 0.99 against the model hedge associated with α = 0.5.
- Transaction costs: The deep-hedging algorithm numerically verifies the proportional transaction-cost relation with rate 2/3 in both Black–Scholes and a two-instrument Heston model.For the tested transaction-cost range, log(pε − p0) = 2/3 log(ε) + C is reported to hold.
- High-dimensional example: A high-dimensional example indicates roughly comparable, close-to-optimal approximation quality and demonstrates potential for higher-dimensional hedging.The comparison uses five models and assesses the neural-network strategy against a bound based on variance-optimal strategies.
6. Disclaimer
The material is informational and not intended as research, advice, recommendation, offer, solicitation, or a basis for evaluating transactions.
- The material states that opinions and estimates reflect the authors’ judgment as of the stated date and may change without notice.
- The material is not a J.P. Morgan Research Department product and was not prepared under legal requirements promoting research independence.
- It is not intended as research, a recommendation, advice, an offer, or solicitation for purchasing or selling financial products or services.
- Readers should consult their own advisors regarding legal, tax, accounting, and other aspects, including suitability implications.
- J.P. Morgan disclaims responsibility for the material’s quality, accuracy, completeness, or any reliance on or use of it.
- Important disclosures are available at www.jpmorgan.com/disclosures.