[Paper] Long-only cryptocurrency portfolio management by ranking the assets: a neural network approach

Source: arXiv - 2512.08124v1

Overview

Zijiang Yang introduces a fresh take on crypto‑asset allocation: instead of forecasting the price of each coin in isolation, a neural network learns to rank the expected returns of a basket of cryptocurrencies and allocates capital accordingly. Tested on daily data spanning the full 2020‑2023 market cycle, the approach delivers a Sharpe ratio above 1 and an annualized return of ~64 %, outperforming several established baselines.

Key Contributions

• Cross‑sectional ranking model – First work to predict a relative return order among many crypto assets rather than absolute price moves.
• Long‑only, rank‑based weighting scheme – Translates predicted ranks into portfolio weights without short‑selling, keeping the strategy realistic for most traders.
• Robust back‑testing – Extensive evaluation on 3.5 years of daily market data covering bull, bear, and sideways periods, with sensitivity analysis to transaction costs.
• Performance edge – Achieves a Sharpe ratio of 1.01 and 64.26 % annualized return, beating classic momentum, mean‑variance, and reinforcement‑learning baselines.

Methodology

1. Data preparation – Daily OHLCV (open, high, low, close, volume) for a curated set of major cryptocurrencies (e.g., BTC, ETH, BNB, SOL, etc.) from May 2020 to Nov 2023.
2. Feature engineering – For each asset and each day, a short window of technical indicators (moving averages, RSI, volatility, volume‑scaled returns) is assembled into a fixed‑size feature vector.
3. Neural ranking network – A feed‑forward network (or lightweight transformer) takes the concatenated feature vectors of all assets at time t and outputs a score per asset. The scores are sorted to produce a predicted ranking of next‑day returns.
4. Weight construction – The top‑k ranked assets receive positive weights proportional to their rank (e.g., linear or exponential decay), while the rest get zero weight. The portfolio is re‑balanced daily, respecting a long‑only constraint.
5. Training objective – A pairwise ranking loss (e.g., hinge loss) encourages the network to assign higher scores to assets that actually outperform others in the next period.
6. Evaluation – Simulated trading with realistic slippage and varying transaction fees (0 %–0.5 %) to test robustness.

Results & Findings

• The model consistently outperformed across all market regimes (bull, bear, flat).
• Performance degradation was modest when transaction fees rose to 0.3 % per trade, confirming the strategy’s cost‑efficiency.
• Ablation tests showed that using the full cross‑sectional input (all assets together) was crucial; training separate models per coin reduced Sharpe to ~0.6.

Practical Implications

• Portfolio construction for crypto funds – The ranking‑based allocation can be integrated into existing execution pipelines with minimal changes; it only requires daily feature updates and a forward pass through a modest‑size neural net.
• Risk‑adjusted returns – Higher Sharpe and lower drawdown make the method attractive for risk‑averse institutional investors who prefer long‑only exposure.
• Scalability – Because the model works on a fixed set of assets, adding new tokens is as simple as extending the feature matrix; the same network can be retrained or fine‑tuned without redesigning the whole system.
• Automation – The daily rebalancing cadence aligns with typical crypto exchange APIs, enabling fully automated bots that react to market‑wide shifts rather than isolated price spikes.

Limitations & Future Work

• Data horizon – The study stops at November 2023; the rapidly evolving crypto landscape (new DeFi tokens, regulatory shocks) may affect out‑of‑sample performance.
• Model simplicity – A relatively shallow feed‑forward architecture was used for interpretability; deeper or attention‑based models could capture richer temporal dynamics.
• Long‑only constraint – While realistic for many traders, the approach ignores potential gains from short positions or derivatives (futures, options). Extending the ranking to include relative short signals is an open avenue.
• Transaction cost model – The back‑test assumes uniform fees; real‑world slippage varies across exchanges and liquidity tiers. Future work could incorporate order‑book simulations for more precise cost estimates.

Bottom line: By shifting the focus from predicting absolute price moves to ranking expected returns across a crypto basket, Yang’s neural‑ranking framework delivers a robust, high‑Sharpe, long‑only strategy that can be readily adopted by developers building automated crypto‑asset managers.

Authors

• Zijiang Yang

Paper Information

• arXiv ID: 2512.08124v1
• Categories: cs.LG, cs.AI, cs.NE
• Published: December 9, 2025
• PDF: Download PDF