Paper: arXiv 2402.06635
Authors: Bryan Kelly, Boris Kuznetsov, Semyon Malamud, Teng Andrea Xu, Yuan Zhang
Abstract
We open up the black box behind Deep Learning for portfolio optimization and prove that a sufficiently wide and arbitrarily deep neural network (DNN) trained to maximize the Sharpe ratio of the Stochastic Discount Factor (SDF) is equivalent to a large factor model (LFM): A linear factor pricing model that uses many non-linear characteristics. The nature of these characteristics depends on the architecture of the DNN in an explicit, tractable fashion. This makes it possible to derive end-to-end trained DNN-based SDFs in closed form for the first time. We evaluate LFMs empirically and show how various architectural choices impact SDF performance. We document the virtue of depth complexity: With enough data, the out-of-sample performance of DNN-SDF is increasing in the NN depth, saturating at huge depths of around 100 hidden layers.
Complexity vs Empirical Score
- Math Complexity: 9.0/10
- Empirical Rigor: 6.0/10 (capped from 8.0: transaction costs not modelled; see penalty flags)
- Quadrant: Holy Grail — high math complexity, high empirical rigor (effective empirical rigor capped from 8.0)
Why this score: The paper is mathematically dense, leveraging advanced deep learning theory (Neural Tangent Kernel) and deriving analytical closed-form solutions for DNN-SDFs. Empirically, it conducts extensive out-of-sample tests using large financial datasets, analyzes architectural choices like depth, and documents performance metrics such as Sharpe ratios and alphas.
Research Flowchart
flowchart TD A["Research Goal: Open the Black Box<br>Is a Deep Neural Network SDF<br>equivalent to a Factor Model?"] --> B["Methodology: Theoretical Construction"] B --> C["Data: Asset Returns<br>and Characteristics"] C --> D["Computation: End-to-End Training<br>Maximizing Sharpe Ratio of SDF"] D --> E["Key Finding 1: DNN SDF =<br>Large Factor Model"] D --> F["Key Finding 2: Depth Matters<br>Performance increases with depth<br>up to ~100 hidden layers"]
Evidence Card
What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card
Proposes
The paper proposes a deep neural network framework that decomposes the stochastic discount factor into nonlinear characteristic discovery and a pricing rule captured by the Portfolio Tangent Kernel (PTK).
Source sentence
We show that a deep neural network (DNN) trained to construct a stochastic discount factor (SDF) admits a sharp additive decomposition that separates nonlinear characteristic discovery from the pricing rule that aggregates them.
Assumes 3
- Factor returns follow Ft+1 = µP + Σ1/2 P Xt with i.i.d. components and finite eighth moments. model
Source sentence
Suppose Ft+1 = µP + Σ1/2 P Xt, where Xt has i.i.d. components with E[X8+ε i,t ] < ∞.
- After a large number of gradient‑descent steps the PTK stabilizes and is non‑degenerate. model
Source sentence
Suppose that, after s∗gradient descent steps, the PTK stabilizes in the sense that
- The data generating process is stationary enough that rolling‑window estimation of the SDF is valid. data
Source sentence
We use these normalized factor returns to estimate a candidate stochastic discount factor as the return on a Markowitz portfolio computed over a rolling window of T months with ridge penalty z.
Data
- JKP characteristics dataset fundamentals public
Source sentence
We use monthly data on 153 stock characteristics for U.S. equities over the period 1963–2024 from (Jensen, Kelly, and Pedersen, 2023, hereafter JKP). This open‑source dataset collects a comprehensive set of firm‑level characteristics
- Random‑feature ReLU benchmark benchmark public
Source sentence
As a benchmark, we construct random‑feature‑based SDFs following Didisheim et al. (2024). Specifically, we define St = ReLU(Xtγ) ∈RNt×P, where γ = (γj,k) ∈Rd×P with P = 25,000 and γj,k ∼N(0, 1) independently.
Costs
Unclear - sources disagree
Finds 5
- The PTK‑SDF achieves substantially higher Sharpe ratios than both the trained DNN and the random‑feature ReLU benchmark across all size groups. performance
Source sentence
Figure 4 confirms this prediction. The PTK‑SDF delivers substantially higher Sharpe ratios than both benchmarks across all size groups.
- A depth‑2, width‑256 DNN attains a Sharpe ratio of 3.8, slightly exceeding the ReLU benchmark’s 3.7. performance
Source sentence
For example, the depth‑2, width‑256 network attains a Sharpe ratio of 3.8, slightly exceeding the ReLU benchmark (3.7).
- Moderately complex DNNs (≈100,000 parameters) discover priced nonlinear characteristics not spanned by random features, yielding a significant positive alpha in spanning tests. statistical
Source sentence
The results strongly support the feature‑learning hypothesis. For moderately complex deep architectures
- Shallow networks (depth = 1) perform poorly and substantially underperform the random‑feature benchmark. negative
Source sentence
Shallow networks (D = 1) perform poorly and substantially underperform the random feature benchmark.
- When the PTK is used, it isolates the economically relevant component of feature learning and translates it into a disciplined pricing rule, yielding tradable SDFs tightly linked to long‑horizon consumption risk. theoretical
Source sentence
Overall, PTK isolates the economically relevant component of feature learning and translates it into a disciplined pricing rule, yielding tradable SDFs that are tightly linked to long‑horizon consumption risk.
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 39 related papers). Coverage is partial, so this is neither replication nor refutation.
2 papers in the archive
Code
No repository by the authors found in the paper.
Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.