平行運算用於即時套利策略交易系統

Abstract

平行運算(Parallel Processing)通常用來解決資料龐大或是計算過程十分複雜的難題。當資料之間的相依性(data dependency)較弱時，平行計算可透過將複雜的難題拆解成許多的小問題，再分配給不同的平行計算單元同時處理，以提升計算效率，減少整體計算時間。 本論文著眼於一個在平行運算中較罕見的應用，提升一個計算複雜度不高，但在即時系統(Real-Time)中計算速度快慢卻有決定性影響的問題—在期貨及選擇權市場上尋找套利機會。在這種高度競爭的市場，套利機會稍縱即逝，因此各券商都在彼此競爭，誰能夠以最快速度送出正確的套利策略單，就會在這場速度競賽中勝出，而比較慢送出的套利單則幾乎沒有成交獲利的機會。

本研究利用平行運算可提升計算速度的特性，建構出一個高效能的衍生性金融商品的套利交易系統。首先建構一個虛擬交易所，藉由台灣期貨交易所提供歷史委託單來模擬真實世界的期貨和選擇權市場，並且修改兩個套利策略，分別是「Convexity of Option Price」(詳見Robert C. Merton(1973))和「Put-Call-Future Parity」(詳見Tucker(1991))。這兩個策略討論的是選擇權價格與期貨價格之間的關係，如果這層關係不存在，立即產生套利機會。 接著在虛擬交易所的環境中加入兩個互相競爭的虛擬交易商，其中一個使用CPU，而另一個使用NVIDIA所開發的平行運算架構CUDA來尋找套利機會，利用GPU所擁有的特殊化平行運算設計，將工作平均地分配到每個計算單元，透過負載均衡以提升平行運算的效能，使得GPU的套利次數和獲利均能打敗CPU，藉此證明套利策略經由平行運算後可大大提升真實市場的獲利率。

Parallel computing is usually applied to solve hard problems with great computational time complexity due to huge amount of data or complicated calculations. The parallel computing can significantly improve the computational efficiency of a big problem with weak data dependencies by spliting the problem into many small and independent problems that can be parallelly solved by different computational units. The overall computational time is thus significantly reduced. This research focuses on a rare application in parallel computing— to improve the computational performance for a light-weight computational problem in a real time environment: finding arbitrage strategies in derivatives markets. Under a competitive environment such as derivatives markets, the speed for searching arbitrage strategies is critical since a late arbitrage order almost fails to get deal. In this thesis, we construct a virtual future exchange to simulate the real world futures and options market. The input orders are the historical data provided by the Taiwan Future Exchange. Two arbitrage strategies being adopted in this paper are modified from the “Convexity of Option Price” (see Robert C. Merton(1973)) and Put-Call-Future Parity (see Tucker(1991)) , which discuss the price relationships between the futures and options. Arbitrage opportunities are found if these relationships are violated. We insert two virtual traders, one use CPU, another one use CUDA, a parallel computing architecture developed by NVIDA, to find the arbitrage opportunities. The GPU can find more arbitrage opportunities and make more profit than the CPU by equally splitting the workload to achieve load balance. We show that finding arbitrage opportunities with parallel computing can greatly enhance profitability in real world financial market.