A fuzzy time series-based neural network approach to option price forecasting Speaker: Prof. Yungho Leu Authors: Yungho Leu, Chien-Pang Lee, Chen-Chia Hung Department of Information Management, National Taiwan University of Science and Technology
Introduction Main idea Fuzzy Time Series The FTSNN Method Option Price Forecasting using FTSNN Results and Performance Conclusion Outline
Option is an important tool for risk management. The premium, also called the price, of an option is determined by many factors. Introduction
The well-known Black-Scholes model (B-S model) was introduced in 1973 to forecast option price. Many limitations limit the use of the B-S model. We propose a hybrid model, FTSNN, that combines fuzzy time series and neural networks to predict option price. Introduction
In FTSNN, the fuzzy time series is used to select training data set and the neural network is used to build the prediction model. We use FTSNN to predict the option price of TXO. “Taiwan Stock Exchange Stock Price Index Options” Introduction
X3X3 X7X7 X5X5 XtXt X1X1 X5X5 X3X3 X t-2 X2X2 X6X6 X4X4 X t-1 X4X4 X8X8 X6X6 ?.... X t-3 X t-5 X t-4 X t-2 X4X4 X2X2 X3X3 X5X5 X t-4 X t-3 X t-1 X t-3 X t-2 XtXt X3X3 X1X1 X2X2 X4X4 X5X5 X3X3 X4X4 X6X6 X t-4 X t-3 X t-1 X t-3 X t-2 XtXt Historical database To predict the next day X t X t-1 X t-3 X t-2 XtXt X t-3 X t-5 X t-4 X t-2 X4X4 X2X2 X3X3 X5X5 X5X5 X3X3 X4X4 X6X6 X t-4 X t-3 X t-1 X t-3 X t-2 XtXt X t-3 X t-5 X t-4 X t-2 Similarly segments Use similar segments to train the prediction model Find similarly segments
X4X4 X2X2 X3X3 X5X5 X5X5 X3X3 X4X4 X6X6 X t-2 X t-4 X t-3 X t-1 X t-3 X t-2 XtXt X t-3 X t-5 X t-4 X t-2 Similarly segments X4X4 X2X2 X3X3 X5X5 X5X5 X3X3 X4X4 X6X6 X t-2 X t-4 X t-3 X t-1 X t-3 X t-2 XtXt X t-3 X t-5 X t-4 X t-2 Using RBFNN to train a prediction model
XtXt X t-2 X t-1 ? To predict the next day X t+1
If F(t) is caused by F(t-1), F(t-2),…,and F(t-n), F(t) is called a one-factor n-order fuzzy time series, and is denoted by F(t-n),…, F(t-2), F(t-1)→F(t).
If F 1 (t) is caused by (F 1 (t-1), F 2 (t-1)), (F 1 (t-2), F 2 (t-2)),…, (F 1 (t-n), F 2 (t-n)), F 1 (t) is called a two-factor n-order fuzzy time series, which is denoted by (F 1 (t-n), F 2 (t-n)),…, (F 1 (t-2), F 2 (t-2)), (F 1 (t-1), F 2 (t-1))→F 1 (t).
Fuzzy Logic Relationship (FLR) Let F 1 (t)=X t and F 2 (t)= Y t, where X t and Y t are fuzzy variables whose values are possible fuzzy sets of the first factor and the second factor, respectively, on day t. Then, a two-factor n-order fuzzy logic relationship (FLR) can be expressed as: (X t-n, Y t-n ), …, (X t-2, Y t-2 ), (X t-1, Y t-1 )→X t, where (X t-n, Y t-n ), …, (X t-2, Y t-2 ) and (X t-1, Y t-1 ), are referred to as the left-hand side (LHS) of the relationship, and X t is referred to as the right-hand side (RHS) of the relationship..
The universe of discourse of the first factor is defined as U= [D min -D 1, D max +D 2 ], where D min and D max are the minimum and maximum of the first factor, respectively; D 1 and D 2 are two positive real numbers to divide the universe of discourse into n equal length intervals. The universe of discourse of the second factor is defined as V= [V min -V 1, V max +V 2 ], where V min and V max are the minimum and maximum of the second factor, respectively; V 1 and V 2 are two positive real numbers used to divide the universe of discourse of the second factor into m equal length intervals.
FTSNN Method Step 2: Define Linguistic terms Linguistic terms A i, 1 ≤ i ≤ n, are defined as fuzzy sets on the intervals of the first factor.
FTSNN Method Step 2: Define Linguistic terms linguistic term B j, 1 ≤ j ≤ m, is defined as a fuzzy set on the intervals of the second factor
For the historical data on day i, let X i-n, Y i-n denote the fuzzy set of F 1 (i-n) and F 2 (i-n) of the fuzzy time series. Let X i denotes the fuzzy set of F 1 (i). The FLRs database on day i can be represented as follows: (X i-n, Y i-n ), …, (X i-2, Y i-2 ), (X i-1, Y i-1 )→X i.
The LHS of the FLR on day t can be represented as follows: (X t-n, Y t-n ), …, (X t-2, Y t-2 ), (X t-1, Y t-1 ).
In the above formulae, IX t-n and IY t-n are the subscripts of the fuzzy terms of the first factor and the second factor, respectively, of the LHS of day t’s FLR. Similarly, RX i-n and RY i-n are subscripts of the first factor and the second factor, respectively, of the LHS of day i’s FLR.
Step 3(e) Model Selection FTSNN uses similar FLRs to build a neural network model. Similar FLRs imply similar trends in the historical data. How long is the trend ? We set the order (length) to be 1,2, …, 5 to build five different prediction models. Then, we choose the best one.
we use the prediction accuracy on day t-1 as the model selection criterion. Error function =| Forecasted RHS - Testing RHS | The forecasted RHS denotes the subscript of the forecasted fuzzy term on day t-1, and the testing RHS denotes the actual subscript of the fuzzy term of the RHS on day t-1
We feed the LHS of the FLR on the predicting day into the neural network to get the forecasted subscript of the RHS on the predicting day. We use weighted average as the defuzzification method. Map the subscript to a value.
where M[k] denotes the midpoint value of the fuzzy term k.
To forecast the price of “Taiwan Stock Exchange Stock Price Index Option (TXO)”. We choose closing price of TXO as first factor and “Taiwan Stock Exchange Capitalization Weighted Stock Index (TAIEX)” as second factor.
Then, we select top five similar FLRs from the FLRs database. In this example, FLR 8, FLR 5, FLR 7, FLR 6, FLR 4 are selected.
A 28 B 117 A 25 B 117 A 30 FLR 8 A 36 B 118 A 37 B 118 A 39 FLR 5 A 39 B 119 A 28 B 117 A 25 FLR 7 A 37 B 118 A 39 B 119 A 28 FLR 6 A 42 B 119 A 36 B 118 A 37 FLR 4 A 28 B 117 A 25 B 117 A 30 A 36 B 118 A 37 B 118 A 39 B 119 A 28 B 117 A 25 Training the prediction model
FLR 8 A 28 B 117 A 25 B 117 A 30 Testing the order of FTSNN A 27 Forecasted fuzzy term A 28 B 117 A 25 B 117 Error =| 27 - 30 | =3 R code
assume that a 2-order neural network model is selected, and the forecasted subscript is 35 on day 11. Substituting 345, 355 and 365 for M[34], M[35], and M[36], respectively. Note that the actual option price on day 11 is 360 in this example.
The dataset of this paper are the daily transaction data of TXO and TAIEX from January 3, 2005 to December 29, Our dataset comprises 30 different strike price from 5,200 to 8,200 and 12 different expiration dates from January 2005 to December 2006.
Two different performance measures, mean absolute error (MAE) and root mean square error (RMSE), are used to measure the forecasting accuracy of FTSNN. where A t and P t denote the real option price and the forecasting option price on day t, respectively.
FTSNN combines a fuzzy time series model and an NN Fuzzy time series model selects training examples for a RBF NN to build the prediction model. The performance of FTSNN is better than the existing models.
Thank you for your attentions
Back