Paper: arXiv 2410.23296
Authors: Ísak Pétursson, María Óskarsdóttir
Abstract
We present a novel approach for predicting the distribution of asset returns using a quantile-based method with Long Short-Term Memory (LSTM) networks. Our model is designed in two stages: the first focuses on predicting the quantiles of normalized asset returns using asset-specific features, while the second stage incorporates market data to adjust these predictions for broader economic conditions. This results in a generalized model that can be applied across various asset classes, including commodities, cryptocurrencies, as well as synthetic datasets. The predicted quantiles are then converted into full probability distributions through kernel density estimation, allowing for more precise return distribution predictions and inferencing. The LSTM model significantly outperforms a linear quantile regression baseline by 98% and a dense neural network model by over 50%, showcasing its ability to capture complex patterns in financial return distributions across both synthetic and real-world data. By using exclusively asset-class-neutral features, our model achieves robust, generalizable results.
Complexity vs Empirical Score
- Math Complexity: 7.0/10
- Empirical Rigor: 8.5/10
- Quadrant: Holy Grail — high math complexity, high empirical rigor
Why this score: The paper presents advanced mathematical concepts including quantile regression, kernel density estimation, and statistical metrics like Wasserstein distance and CRPS, demonstrating high mathematical sophistication. It is highly empirical, utilizing real-world and synthetic datasets, public code repositories, backtest-ready evaluation metrics (VaR, CRPS), and rigorous out-of-sample testing across multiple asset classes.
Research Flowchart
flowchart TD
A["Research Goal: Predict<br>Asset Return Distributions"] --> B["Method: Two-Stage<br>LSTM Quantile Regression"]
B --> C["Stage 1: Predict Asset-Specific<br>Quantiles using Features"]
C --> D["Stage 2: Adjust Predictions<br>with Market Data"]
D --> E["Compute: Kernel Density<br>Estimation (KDE)"]
E --> F["Outcomes: Full Probability<br>Distribution Prediction"]
F --> G{"Performance Results"}
G --> H["98% improvement vs Linear Quantile Regression"]
G --> I[">50% improvement vs Dense NN"]
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 generalizable distribution estimation model that uses only asset‑neutral features, predicts multiple periods of return quantiles, and combines them with a hybrid Gaussian‑quantile approach.
Source sentence
Proposes a generalizable distribution estimation model that: – Uses only asset-neutral features. – Allows the consideration of multiple periods of return.
Assumes 6
- Asset‑neutral features are sufficient to capture return dynamics across all asset classes. data
Source sentence
ensuring that each characteristic is asset-class neutral, which makes the approach applicable across different asset classes.
- Quantile regression can be converted into a full probability distribution via smoothed density estimation. model
Source sentence
we convert these quantiles into a full distribution using a smoothed density estimation function.
- Synthetic returns are scaled by randomly sampled means and standard deviations to mimic realistic market conditions. statistical
Source sentence
For each distribution, the synthetic returns rsynthetic are scaled by the randomly sampled mean ¨µ ∼Uniform(−0.001, 0.001) and standard deviation ¨σ ∼Uniform(0.01, 0.03).
- All market data are normalized using rolling Z‑scores with a window length treated as a hyperparameter. data
Source sentence
All data are normalized using rolling Z-scores, with the window length set as a hyperparameter to be optimized during model training.
Show 2 more
- The model is trained and evaluated on daily data spanning from February 3, 2000 to January 1, 2024. market
Source sentence
The data spans from February 3, 2000 up to January 1, 2024, sourced from the Yahoo Finance API.
- Quantile loss is an appropriate metric for training and evaluating the model across future time steps. statistical
Source sentence
training the model to minimize the quantile loss for each future time step within the horizon.
Data
- Yahoo Finance API price data market-prices public
Source sentence
The data spans from February 3, 2000 up to January 1, 2024, sourced from the Yahoo Finance API.
- Synthetic datasets per distribution synthetic public
Source sentence
We generate 10 synthetic datasets per distribution, each consisting of 1000 samples, to assess the model’s robustness under these simulated conditions.
Costs
Not applicable
Finds 5
- The qLSTM model achieves the lowest total quantile loss for real‑world data (0.4065) and overall (0.3840), outperforming baselines. performance
Source sentence
LSTM achieves the lowest total quantile loss for real-world data (0.4065), while the qHybrid model marginally outperforms it for synthetic datasets (0.2987 vs. 0.3503). When considering all datasets, the LSTM model maintains the lowest overall total quantile loss (0.3840).
- The qHybrid model attains the lowest Wasserstein distance for most asset classes, especially cryptocurrencies (0.0213) and commodities (0.0072). performance
Source sentence
classes, particularly for cryptocurrencies (0.0213) and commodities (0.0072).
- At the extreme tail (0.00075 level), qLSTM consistently records the lowest VaR across most asset classes. performance
Source sentence
At the extreme tail (0.00075 level), the qLSTM model consistently achieves the lowest VaR across most asset classes, including FX (0.0008), commodities (0.0028), Euro Stoxx 50 (0.0031), S&P 500 (0.0024), and Nikkei 225 (0.0018).
- The LSTM model outperforms qHybrid on simpler synthetic distributions (Normal, Uniform) in terms of quantile loss. performance
Source sentence
However, for simpler distributions such as Normal and Uniform, LSTM outperforms the qHybrid model, indicating its ability to generalize across different statistical patterns.
- The qLSTM model achieves the lowest RMSE for mean estimation across all asset classes, while the Constant model has the highest RMSE. statistical
Source sentence
For the mean µ, the qLSTM model consistently achieves the lowest RMSE in all asset classes, highlighting its superior ability to accurately estimate the
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 28 related papers). Coverage is partial, so this is neither replication nor refutation.
Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.