期货作为现代金融市场的重要组成成分之一，持续发挥着套期保值、价格发现等作用。在国外的成熟期货市场中，近半数交易均为套利交易，而期货的套利主要包括跨期套利、跨市套利和跨品种套利。本文对铁矿石、焦炭和螺纹钢期货的黑色金属产业链进行了跨品种套利的研究，以期为钢铁企业降低市场价格波动风险，助力企业和行业的稳定发展。 在对铁矿石、焦炭和螺纹钢期货进行基本面分析后，本文在Wind选取2013年11月1日至2016年11月1日的铁矿石、焦炭和螺纹钢期货主连合约的日收盘数据做训练集，2016年11月2日至2017年11月1日的数据做测试集来划分样本，来对铁矿石、焦炭和螺纹钢期货进行跨品种套利研究。 本文根据经典统计套利原理，对训练集数据进行ADF检验、相关性分析和协整检验后，将铁矿石、焦炭和螺纹钢期货的对数价格数据建立了长期均衡的协整回归方程和短期波动的误差修正模型，结果显示铁矿石、焦炭和螺纹钢期货三者的对数价格之间确实在统计意义上存在长期稳定的均衡关系，并在发生短期波动时能够负反馈调节。 根据经典协整回归模型进行套利交易后表明策略收益良好，在此基础上，本文从是否考虑加仓、BP神经网络和Elman神经网络三个角度来对协整方程的经典套利策略进行了优化。训练集的实证研究表明：是否考虑加仓的经典回归模型和两种神经网络在总利润的规模上并没有太大的差异，但神经网络套利的总亏损都远远小于是否考虑加仓的回归模型，且能够很大程度上防止资金回撤，有效减少交易成本，且多为短期交易，能较大程度上防控长期持仓带来的保证金波动风险；结合Sharpe ratio、套利效果、实际应用的便捷性和模型复杂度，本文选取BP神经网络作为经典均值回归的最优优化模型。 确定了经典协整回归模型的最优优化模型后，本文对测试集数据进行实证，结果显示：测试集的BP神经网络套利胜率一般，同样有着多次短期交易的特点，使得整个交易过程有平均持仓时间短、交易频率高、年化收益率高的特点，表现合理，具有实践的可行性。 本文的创新点主要在于，首先，现有期货跨品种套利的成果主要集中在两品种的协整回归模型，本文则从产业链的角度出发，建立了多品种的跨品种套利；其次，铁矿石期货于2013年推出，其套利研究成果不多，本文黑色金属产业链跨品种套利研究是对铁矿石期货套利的有益补充；最后，目前国内已有的套利策略研究大多聚焦于均值回归模型，本文在经典回归模型的基础上，从是否考虑加仓、BP神经网络和Elman神经网络三个角度来进行优化，对比协整回归模型的利润价差套利与神经网络模型的利润拟合预测套利的优劣，运用四种方法对黑色金属产业链的跨品种套利策略进行研究，为进一步探索黑色金属产业链跨品种套利提供一定的理论支持。

As one of the most important components of the modern financial market, futures has received increasing attention. With the improvement of China’s international status, the futures market has been improved gradually, and play an irreplaceable role in hedging and price discovery. In foreign mature futures markets, nearly half of the transactions are arbitrage transactions, and the arbitrage of futures mainly includes intertemporal arbitrage, cross-market arbitrage and cross-variety arbitrage. This paper consider to research on cross-variety arbitrage about black industrial chain of the futures of iron ore, coke, and rebar, in order to reduce the risk of market price fluctuations for iron and steel companies and help the industries and companies develop stably. After the fundamental analysis of iron ore, coke, and rebar futures, this paper selects the daily closing data of iron ore, coke, and rebar futures contracts to conduct cross-variety arbitrage research. Our train set is from November 1, 2013 to November 1, 2016 and test set is from November 2nd, 2016 to November 1st, 2017. Based on the classical statistical arbitrage principle, this paper conducts ADF unit root test, correlation analysis and cointegration test on the train set, and established a long-term equilibrium cointegration regression equation for the logarithmic price data of iron ore, coke, and rebar futures. And we also built a short-term fluctuations in the error correction model. The result shows that there are long-term, stable relations between logarithmic prices of iron ore, coke, and rebar futures, can be adjusted by negative feedback in the event of short-term fluctuations. Based on the arbitrage strategy based on the classical co-integration regression model, it shows that the return is handsome. On this basis, this paper optimizes the classical arbitrage strategy of cointegration equations by considering whether adding warehousing, BP neural network or Elman neural network. The empirical study of the train set shows that there are not much difference between whether or not to consider adding positon in classical regression model and the two neural network in the scale of total profit, but the total loss of neural network arbitrage is much smaller than those two. Moreover, neural network can greatly prevent the withdrawal of funds and effectively reduce transaction costs, most of them are short-term transactions, which can prevent and control the risk of margin fluctuations caused by long-term positions. Combining Sharpe ratio, arbitrage effects, convenience of practical application, and model complexity, I select BP neural network as the best optimal optimization model. After selecting the best optimization model of the classical statistical arbitrage model, this paper empirically calculates the test set. The result shows that the arbitrarily winning ratio of the BP neural network is normal. It also has the characteristics of multiple short-term transactions, which shorten the entire average transaction greatly. With the characteristics of short position, high transaction frequency and high-annualized rate of return, the performance of the model on test data is reasonable and has practical feasibility. The innovation of this paper is, first, the results of cross-variety arbitrage of existing futures are mainly concentrated on the two varieties, but this article establishes a multi-variety of cross-variety arbitrage of the industrial chain. Second, swap futures came into the market since 2013, there are not so many researches on it. This article is a useful supplement to the swap futures research. Last but not the least, most existing research focuses on the cointegration regression model, this paper optimizes the classical regression model from whether or not to consider adding positon, BP and Elman neural network, comparing the profit of the cointegration regression model and neural network, using four methods to study on the ferrous metal industrial chain's cross-variety arbitrage strategy, to provide further theoretical support for it.