Deep Learning for Financial Time Series: A Large-Scale Benchmark of Risk-Adjusted Performance

This paper presents a large-scale benchmark of deep learning architectures for financial time series prediction and position sizing, demonstrating that hybrid models like VSN with LSTM and xLSTM outperform linear and generic deep learning baselines by achieving superior risk-adjusted returns, robustness to transaction costs, and improved downside protection on a diverse daily futures dataset from 2010 to 2025.

The Gist

Imagine you are trying to predict the weather to decide whether to carry an umbrella, but instead of rain, you are predicting whether the stock market will go up or down. And instead of a simple "yes/no," you have to decide how much money to bet on it.

This paper is like a massive, high-stakes cooking competition where 15 different chefs (AI models) are asked to cook up the best investment strategy using the same set of ingredients (15 years of financial data from stocks, bonds, and commodities).

Here is the breakdown of what happened, using simple analogies:

1. The Goal: The "Sharpe Ratio" Scorecard

In the world of investing, you don't just want to make money; you want to make money without losing your shirt when things go wrong.

• The Analogy: Imagine two drivers. Driver A drives 100 mph but swerves wildly, barely missing trees. Driver B drives 60 mph but stays perfectly in the lane.
• The Paper's Metric: They used a score called the Sharpe Ratio. It measures how much "smoothness" (safety) you get for every mile of speed (profit). The goal wasn't just to be the fastest car; it was to be the smoothest, most reliable ride.

2. The Contestants: The "Chefs"

The researchers tested five different types of "chefs" (AI architectures) to see who could handle the chaotic kitchen of the financial market:

• The Old School (Linear Models): These are like chefs who only know how to make a simple sandwich. They assume the future is just a straight line from the past.
  • Result: They got hungry. They sometimes worked well when the market was calm, but they failed miserably when things got crazy. They couldn't handle the chaos.
• The Transformers (The "Attention" Chefs): These are the new, trendy chefs who look at everything at once. They are great at reading long books but sometimes get overwhelmed by the noise in a busy kitchen.
  • Result: They were okay, but often got distracted by the noise. They struggled to find the signal in the static.
• The State-Space Models (The "Memory" Chefs): These are like chefs who have a perfect memory of every ingredient they've ever touched, but they try to process it all in a straight line.
  • Result: They were fast and efficient, but they lacked the "gut feeling" needed to adapt to sudden market changes.
• The Recurrent Networks (The "Storytellers"): These chefs read the story of the market one word at a time, remembering the plot as they go.
  • Result: They did very well, especially the new, upgraded versions.

3. The Winners: The "Hybrid" Masters

The paper found that the best chefs weren't the ones using just one trick. The winners were Hybrids—models that combined different superpowers.

• The Champion (VLSTM): This model is like a Smart Sous-Chef.

  • How it works: It first uses a filter (Variable Selection Network) to ignore the noisy, useless ingredients (like a bad smell in the kitchen) and focus only on the fresh, important ones. Then, it passes that clean list to a master storyteller (LSTM) who remembers the long-term trends.
  • Why it won: It had the highest "Sharpe Ratio." It made the most money for the least amount of stress. It was the most reliable driver.

• The Runner-Up (xLSTM): This is the Resilient Veteran.

  • How it works: It's an upgraded version of the old storyteller. It has a better memory system that doesn't forget important details even if they happened a long time ago.
  • Why it's special: It was the most robust. If you had to pay a "fee" every time you changed your mind (transaction costs), this chef lost the least amount of money. It didn't over-trade; it waited for the right moment.

4. The Big Lessons (The "Takeaways")

• Simple isn't enough: You can't just draw a straight line through the stock market. The market is too messy, too noisy, and changes its mind too often. You need a model that understands context.
• Noise is the enemy: Financial data is like a radio station with a lot of static. The winning models were the ones best at turning down the static (noise) and turning up the music (the real signal).
• Don't just look at the profit: A model that makes 100% profit but crashes your car (huge losses) is a bad model. The best models were the ones that kept you safe during the worst storms (downside risk).
• The "Cost" of Trading: The paper showed that some models trade so much they get eaten alive by fees. The best models were efficient—they didn't trade just for the sake of trading.

Summary

Think of this paper as a report card for the future of AI in finance. It tells us that the old, simple math is out. The new winners are smart, hybrid systems that can filter out the noise, remember the important parts of the story, and stay calm when the market goes crazy.

The ultimate lesson? In finance, the best investor isn't the one who makes the most money on a lucky day; it's the one who stays in the game the longest by avoiding disaster. The "VLSTM" and "xLSTM" models proved to be the best at keeping the lights on.

Technical Summary

1. Problem Statement

The paper addresses the challenge of applying modern deep learning architectures to financial time-series prediction and position sizing. Unlike standard time-series benchmarks (e.g., weather or electricity demand) which often exhibit strong seasonality and high signal-to-noise ratios, financial data is characterized by:

• Low signal-to-noise ratios: Predictable signals are weak and buried in noise.
• Non-stationarity: Market regimes change over time.
• Heavy-tailed distributions: Returns exhibit extreme outliers and volatility clustering.
• Regime dependence: Models must adapt to shifting market conditions without overfitting to transient patterns.

The core objective is to determine which deep learning architectures can effectively extract economically meaningful signals from this noisy data to maximize risk-adjusted returns (Sharpe Ratio) rather than just minimizing prediction error (MSE).

2. Methodology

2.1 Dataset and Scope

• Data: A comprehensive dataset spanning 2010–2025 covering five asset classes: Commodities, Equity Indices, Bonds, Foreign Exchange (FX), and Energy.
• Features: End-of-day prices, returns, and technical indicators (e.g., momentum, volatility-normalized MACD).
• Preprocessing: Continuous futures contracts were constructed using back-adjusted ratios. Volatility was estimated using an Exponentially Weighted Moving Average (EWMA).

2.2 Experimental Framework

The authors propose an end-to-end portfolio optimization pipeline (Figure 1):

1. Input: A lookback window of historical features ($X_t$).
2. Model Architecture: A sequence model ($g_\phi$) extracts a temporal state representation ($h_t$).
3. Projection: A linear layer followed by a tanh activation maps the state to a trading signal $\hat{y}_t \in [-1, 1]$ (representing long/short conviction).
4. Position Sizing: Signals are scaled by a volatility targeting mechanism ($1/\sigma_t$) to equalize risk contributions across assets.
5. Optimization: Unlike traditional models that minimize Mean Squared Error (MSE), this framework directly optimizes the Sharpe Ratio. The loss function is the negative annualized Sharpe Ratio:
   $$L(\theta) = -\frac{\hat{E}[R]}{\sqrt{\hat{Var}[R]}}$$
   This forces the network to learn representations that maximize expected return while penalizing variance.

2.3 Evaluated Architectures

The benchmark covers a wide spectrum of modern and classical models:

• Linear Baselines: AR1x, DLinear, NLinear.
• Transformer-Based: iTransformer, PatchTST, Temporal Fusion Transformer (TFT).
• State-Space Models (SSMs): Mamba, Mamba2.
• Recurrent Networks: LSTM, xLSTM (with exponential gating), Patch sLSTM (PsLSTM).
• Hybrid Models:
  • VLSTM: Variable Selection Network (VSN) + LSTM.
  • VxLSTM: VSN + xLSTM.
  • LPatchTST: LSTM (denoiser) + PatchTST.
  • VSN+Mamba2: Feature selection prior to SSM.

2.4 Evaluation Metrics

The study goes beyond average returns to include:

• Risk-Adjusted Performance: Sharpe Ratio, Information Ratio, CAGR.
• Downside Risk: Maximum Drawdown (Max DD), Calmar Ratio, Worst 3-month Sharpe, CVaR (5%).
• Robustness: Sensitivity to random seed selection (tested with reduced seed aggregation).
• Implementation Friction: Breakeven Transaction Cost ($c^*$) analysis to determine the maximum cost a strategy can endure before becoming unprofitable.
• Computational Efficiency: Parameter counts and asymptotic complexity.

3. Key Contributions

1. Unified Benchmark: The first large-scale comparison of diverse deep learning architectures specifically optimized for Sharpe Ratio rather than prediction error, across 15 years of multi-asset data.
2. Direct Risk Optimization: Demonstrating that optimizing the Sharpe Ratio directly (via differentiable loss) yields superior economic performance compared to standard forecasting approaches.
3. Hybrid Architecture Superiority: Identifying that hybrid models combining feature selection (VSN) with structured temporal representations (LSTM/xLSTM) outperform both pure linear models and generic attention-based models.
4. Robustness Analysis: Providing evidence that top-performing models are not artifacts of favorable random seed selection but possess structural advantages in handling non-stationarity.
5. Transaction Cost Sensitivity: Introducing a rigorous breakeven cost analysis showing that some models (like xLSTM) achieve higher efficiency by reducing turnover, even if their gross Sharpe is slightly lower than others.

4. Key Results

4.1 Performance Rankings

• Top Performer: VLSTM (VSN + LSTM) achieved the highest overall Sharpe Ratio (2.40 over 2010–2025) and the highest CAGR (26.3%). It demonstrated exceptional stability across different market regimes (e.g., 2020–2025).
• Strong Contenders:
  • LPatchTST (LSTM + PatchTST): Sharpe 2.31, excellent downside-adjusted characteristics.
  • TFT: Sharpe 2.27.
  • xLSTM: Sharpe 1.79, but with the highest breakeven transaction cost buffer, indicating superior robustness to trading frictions.
• Underperformers:
  • Linear Models (DLinear, NLinear): Failed to deliver stable performance; Sharpe ratios remained below 1.0.
  • Pure Transformers (PatchTST, iTransformer): Showed high sensitivity to specific years and lower aggregate Sharpe ratios (e.g., PatchTST ~0.76).
  • State-Space Models (Mamba2): Showed heterogeneous behavior; while computationally efficient, they required feature conditioning (VSN) to become competitive (Sharpe ~1.10).

4.2 Risk and Downside Protection

• VLSTM and LPatchTST exhibited moderate drawdowns and high Calmar ratios.
• VxLSTM achieved the lowest maximum drawdown (-11.8%) and highest Calmar ratio (1.64), suggesting a more conservative but stable return profile.
• Standard LSTMs and PatchTST suffered larger drawdowns, indicating higher sensitivity to extreme market movements.

4.3 Robustness and Efficiency

• Seed Robustness: When reducing the experimental budget (from top 10 seeds to top 5), the ranking of models remained stable, confirming that the performance gains are structural, not due to lucky initialization.
• Transaction Costs: xLSTM demonstrated the best trade-off between signal quality and turnover, allowing it to withstand higher transaction costs than VLSTM, despite VLSTM having a higher gross Sharpe.
• Computational Cost: Linear models were cheapest but ineffective. Hybrid models (VLSTM, LPatchTST) had higher parameter counts but offered the best empirical return on investment.

5. Significance and Conclusions

The paper concludes that architectural inductive bias is decisive in financial forecasting.

• Linear dynamics are insufficient: Purely linear models cannot capture the structural features of financial returns.
• Structured representations matter: Models that explicitly learn structured temporal representations (via gating, memory compression, or feature selection) outperform generic attention mechanisms.
• The "Sweet Spot": The most successful architectures (VLSTM, LPatchTST) combine:
  1. Adaptive gating to handle regime shifts.
  2. Representation compression (e.g., VSN) to filter noise.
  3. Stable state evolution to maintain memory of long-range dependencies without overfitting.
• Practical Implication: For real-world deployment, the "best" model depends on the investor's constraints. If transaction costs are high, xLSTM may be preferable due to lower turnover. If maximizing raw risk-adjusted return is the goal, VLSTM is superior.

This benchmark provides a transparent reference point for future research, emphasizing that in finance, robustness, tail-risk control, and implementation feasibility are as critical as raw predictive accuracy.