證券交易所的限價單資料是在極短時間間隔所產生的大量高頻交易資料，和傳統的 股票交易資料有著截然不同的序列結構。本篇論文探討如何利用長短期記憶網絡模型建 立高頻交易資料的預測模型。我們研究長短期記憶模型的架構及高頻交易資料的處理方 式，並揭露了以高頻交易資料建立長短期記憶模型時所面臨的問題。最後以平均平方誤 差及平均絕對誤差為準則建立預測模型並和其他機器學習模型 (支持向量機，決策樹，K 鄰近法，多層感知器) 比較其預測優劣。

Limit orders data of the stock exchange is in a very short time interval produced by a large number of high-frequency data, and the traditional stock trading data sequence with different structure. This paper discusses how to use LSTM model to build the prediction model of high- frequency trading data. We study the structure of long and short term memory model and the processing method of high frequency trading data, and reveal the problems in building long and short term memory model with high frequency trading data. Finally, the average square error and average absolute error were used as the criteria to establish the prediction model, and compared with other machine learning models (support vector machine, decision tree, K neighborhood method, multi-layer perceptron).

論文審定書 i
論文公開授權書 ii
誌謝 iii
摘要 iv
Abstract v

1 Introduction 1
1.1 Motivation...................................... 1
1.2 Goals ........................................ 1
1.3 Report structure................................... 1
2 Data description 3
2.1 HFTdata ...................................... 3
2.2 Data preprocessing ................................. 3
3 Methods and materials 5
3.1 Recurrent Neural Network ............................. 5
3.2 The core idea behind LSTMs............................ 5
3.3 How to learn LSTM................................. 7
3.3.1 Loss function................................ 7
3.3.2 Back propagation through time....................... 7
3.3.3 Parameters update ............................. 8
3.4 Other parameter................................... 10
3.4.1 Activation function............................. 10
3.4.2 Epochs and batch size ........................... 12
3.4.3 Shape of model............................... 12
4 Design of model 15
4.1 Previous experience................................. 15
4.1.1 Pricepredict ................................ 15
4.1.2 Log return predict ............................. 15
4.1.3 Different shape of model.......................... 15
4.1.4 The issues of regularization ........................ 16
4.1.5 Long memory v.s short memory ...................... 16
4.1.6 Multiple variable model .......................... 16
4.2 Data preparation................................... 16
4.2.1 Turn to mean time ............................. 16
4.2.2 Build the sequence that we’re interested in . . . . . . . . . . . . . . . . 16
4.2.3 Sliding window............................... 17
4.3 Determine the sequence patterns .......................... 18
5 Empirical study 19
5.1 Data description................................... 19
5.2 Data processing................................... 19
5.3 Parameter selection ................................. 20
5.4 Example 1: Midquote ............................... 21
5.5 Example 2: Log ratio of supply to demand..................... 24
5.6 Example 3: Bid ask spread ............................. 27
6 Conclusion 30
7 References 31
8 Appendix 32
8.1 Parameter comparison of mid quote(fix sliding window(10,5)) . . . . . . . . . . 32
8.2 Parameter comparison of log ratio of supply to demand . . . . . . . . . . . . . . 33
8.3 Parameter comparison of bid ask spread ...................... 35
8.4 Other tables..................................... 37
8.5 Acf and pacf plot .................................. 38

[1] Bianchi, F. M., Maiorino, E., Kampffmeyer, M. C., Rizzi, A., Jenssen, R. (2017). An overview and comparative analysis of Recurrent Neural Networks for Short Term Load Forecasting. arXiv preprint arXiv: 1705.04378.
[2] Goodfellow, I., Bengio, Y., Courville, A., and Bengio, Y. (2016). Deep learning (Vol. 1). Cambridge: MIT press.
[3] Ge ́ron,A.(2017).Hands-onmachinelearningwithScikit-LearnandTensorFlow:concepts, tools, and techniques to build intelligent systems. O’Reilly Media.
[4] Graves, A. (2012). Supervised Sequence Labelling with Recurrent Neural Networks. Springer, New York.
[5] Zankova, E. (2016). High frequency financial time series prediction: machine learning ap- proach.