Deep Learning for Arbitrage Trading in Crypto Options Market

1. Introduction
Options are derivative instruments whose value is linked to an underlying asset. Compared with spot trading, options markets are more complex and often less liquid, which can leave more short-lived pricing inconsistencies that are not immediately competed away. In cryptocurrency options markets, these inconsistencies can appear as temporary arbitrage windows. However, not every observed window is worth acting on: some are brief microstructure noise (sometimes called toxic flow), while others persist long enough to be meaningfully exploitable. Allocating capital to an opportunity that disappears immediately, or moves against the position before converging, is inefficient.
We have a proprietary dataset consisting of one year of per-second limit order book (LOB) snapshots from Delta Exchange, with timestamps labeled for when such arbitrage windows occur. The goal of this thesis is to use machine learning on these LOB snapshots to extract informative features and predict actionability, that is, whether an observed arbitrage signal is likely to remain open long enough and evolve favorably enough to justify taking it.

2. Problem Statement
The challenge is to use the rich information in the full limit order book (LOB), such as market depth, buy sell imbalance, and short-term order flow pressure, to make two predictions before taking any action:

• Life-span: How long is this arbitrage window likely to remain open?
• Short-term direction: Over the next few seconds, is the opportunity more

likely to improve (move toward closure) or worsen (move further away) before it closes?

3. Thesis Objectives
The student will develop a deep learning model to act as a "Viability Filter" for the trading system:
a. LOB Feature Engineering: Convert raw LOB snapshots into predictive
tensors (e.g., converting price levels/volumes into images for CNNs or sequences for LSTMs).
b. Multi-Task Prediction: Train a model (e.g., DeepLOB or Transformer) that
simultaneously predicts:

• Classification: Is the gap "Tradeable" vs. "Noise"?
• Regression: Expected time-to-closure (in seconds).
• Forecasting: The probability of the gap widening in the next t seconds.

c. Budget Optimization Strategy: Simulate a margin-constrained portfolio
where capital is dynamically allocated only to the highest-confidence
predictions, comparing the PnL against a naive "trade-all" baseline.

4. Technical Stack

• Data: Delta Exchange LOB snapshots + PCP marks (parquet/HDF5),
• Models: PyTorch/TensorFlow,
• Architectures: CNN-LSTM (DeepLOB), Temporal Fusion Transformers,
• Metrics: Precision/Recall on "Tradeable" classification, Sharpe Ratio of the
• resulting strategy.

5. References:
1. Alexander, C., Chen, X., Deng, J., Wang, T. (2024). Arbitrage opportunities and eƯiciency tests in crypto derivatives. Journal of Financial Markets, 71, 100930. Topic: Deep Learning for Arbitrage Trading in Crypto Options Market
2. Felföldi-Szűcs, N., Králik, B., Váradi, K. (2024). Put–call parity in a crypto option market: Evidence from Binance. Finance Research Letters, 61, 104874.
3. Cartea, Á., Jaimungal, S., Penalva, J. (2015). Algorithmic and High-Frequency Trading. Cambridge University Press.
4. Makarov, I., Schoar, A. (2020). Trading and arbitrage in cryptocurrency markets. Journal of Financial Economics, 135(2), 293–319

None