摘要
论文题目：融入投资者关注度的在岸与离岸人民币价差的预测
论文类型：方案策划
专业方向：数据分析
摘 要
伴随着我国经济在过去几十年的强劲表现，国际上人民币的使用也在
迅速增长，而我国在新冠疫情中展现出的优异表现也助推了人民币国际化
的深入。自从我国央行于 2010年 7月正式批准人民币离岸交易以来，离岸
人民币汇率开始形成并得以发展。两种人民币汇率的形成机制不同使得在
岸人民币与离岸人民币之间存在着天然的价差，于是本文从过往研究在、
离岸人民币价差形成的因素中初步筛选出一些影响变量。此外，自耶鲁学
派的行为金融学理论受到广泛关注以来，有关于投资者关注度的量化方式
在学界中不断地推陈出新，市场中天然存在着可以部分勾勒投资者关注度
的经济金融指标，但是为了更客观、更直接揭示投资者的真实关注度，本
文利用《哈佛大学社会心理学词典》、《拉斯韦尔价值词典》中经济学领域
的词汇和谷歌趋势的词汇搜索量共同构建出投资者关注度指标 FEARS。之
后，利用 SHapely Additive exPlanation对各变量的重要性进行排序与筛选，
具体可以发现 3个月的在岸人民币 NDF与离岸人民币 NDF的差值对在、
离岸人民币价差的形成有重大贡献，这一点与远期合约的价格发现功能所
描述的内容不谋而合，此外本文构建出的投资者关注度指标 FEARS对价
差形成的贡献有着不对称性，这一点也与一般逻辑相符合。之后为了实现
更好的预测效果，本文将奇异谱分析运用于汇率数据的分解与重构之中，
以使数据更加平稳，从而更直接地揭示该序列数据中的重要趋势。最后，
从 SHAP模型中选取贡献度最高的 6个变量，并在进行了奇异频谱分析后，
建立复合神经网路模型 CNN-BiGRU以对在、离岸人民币价差进行预测，
并从多个回归模型的评价指标着手，与单一的 CNN模型、 GRU模型、
ARIMA模型进行对比，发现纳入了投资者关注度 FEARS的 CNN-BiGRU
模型的预测能力是较为优秀的。
关键词：投资者关注度；在岸与离岸人民币价差；奇异谱分析

Abstract
Abstract
Along with the strong performance of the Chinese economy over the past de
cades, the international use of the RMB has grown rapidly, and the excellent
performance demonstrated by China in the COVID-19 epidemic has also con
tributed to the deepening internationalization of the RMB. The offshore RM
B exchange rate has begun to take shape and has developed since the People’
s Bank of China officially approved offshore RMB trading in July 2010. The
difference in the formation mechanism of the onshore and offshore RMB ex
change rates results in a natural spread between these two exchange rates, so
this paper initially screens out some of the variables from previous studies o
n the formation of the onshore and offshore RMB spreads . In addition, since
the Yale School's behavioural finance theory has received widesprea d attenti
on, there are new ways of quantifying the attention of investors, which are c
onstantly being developed in the academic circle. There are naturally econo
mic and financial indicators in the market that can partially outline investors'
attention, but in order to reveal the real attention of investors in a more obje
ctively and directly way, this paper uses economic words from《Harvard Dic
tionary of Social Psychology》、《Laswell Value Dictionary》and the search
volume from Google Trends to jointly construct the indicator of the attention
of investors, FEARS. After that, the importance of each variable is ranked a
nd filtered using SHapely Additive explanation (SHAP). Specifically, it can
be found that the difference between the 3-month onshore RMB Non-deliver
able Forwards (NDF) and offshore RMB NDF has a significant contribution t
o the formation of the spread between the onshore and offshore RMB , which
coincides with what is expressed by the price discovery function of the forw
ard contract. Besides, the contribution of FEARS, an indicator of investor s’ a
ttention, to the formation of the spread is asymmetric, which is also consiste
nt with the general logic. The paper then applies singular spectrum analysis
(SSA) to the decomposition and reconstruction of the exchange rate data in o
II

Abstract
rder to make the data smoother and thus reveal the important trends in this se
ries data more directly to achieve better forecasting results. Finally, the six v
ariables with the highest contributions from the SHAP model were selected a
nd, and after the singular spectrum analysis, a composite neural network mo
del CNN-BiGRU was built to forecast the onshore and offshore RMB spread
s. Different evaluation indicators of multiple regression models were calcula
ted and compared with a single CNN model, GRU model and ARIMA model,
and it was found that the forecasting ability of the CNN -BiGRU model, whi
ch incorporates FEARS, the attention of investors, was superior.
Keywords : Attention of Investors; Spread between Onshore and Offshore R
MB; Singular Spectrum Analysis
III
。。。以下略