Financial Time series Forecasting using support vector machines ( Elaborated by Mohamed DHAOUI , 3rd engineering student at Tunisia Polytechnic School ) .

1. Tunisia Polytechnic School
Data Mining project
Presented by
Mohamed DHAOUI
3rd year engineering student
(contact@Mohamed-dhaoui.com) Academic Year : 2015-2016
Financial time series forecasting using
support vector machines

2. 2

3. 


◦
◦
◦
3

4. 4

5. 5

6. 6

7. 7

8. 8

9. 9

10. 10

11. 11

12. 12

13. 13

14. a weight parameter, which needs to be
carefully set
14

15. 15

16. 16

17. 17

18. 18

19. 19

20. 

20

21. 21

22. 22

23. 23

24. 24

25. 25

26. 26

27. 27

28.  Backpropagation, an abbreviation for
"backward propagation of errors", is a
common method of training artificial neural
networks used in conjunction with
an optimasition method such as gradient
descent. The method calculates the gradient
of a loss function with respect to all the
weights in the network and try to update
these weights,
28

29. 29

30.  Algorithm:
 initialize network weight (randomly)
 Do
forEach training example ex
prediction = neural-net-output(network, ex)
actual = teacher-output(ex)
compute error (prediction - actual) at the output units
compute for all weights
update network weights
until all examples classified correctly or another stopping criterion satisfied
 return the network
30

31.  Weights updating
 Δwt = -e* E + α Δwt-1
 =H* δ0
e: learning rate
α :momentun
Wh,o
Hidden
layer
Output
layer
O
H
E= actual-ideal
δ0= -E*f’(o)
δk= f’(h)*Wh,o *δ0
31

32.  Weaknesses
• Gradient descent with backpropagation is not
guaranteed to find the global minimum.
• There is no rule for selecting the best
learning rate and the momentum.
• Slow algorithm that need a computational
resources.
32

33. SVM perfermance
• Too small value for C caused
underfit the training data
while too large a value of C
caused overfit the training
data
33

34. the best prediction performance of the holdout data
is recorded when delta is 25 and C is 78
SVM perfermance
34

35. BP perfermance
• The best prediction performance for the holdout data is
produced when the number of hidden processing elements are
24 and the stopping criteria is 146 400 epochs.
• The prediction performance of the holdout data is 54.7332%
and that of the training data is 58.5217%.
35

36. Comparison
 SVM outperforms BPN and CBR by 3.0981% and
5.852% for the holdout data, respectively
 For the training data, SVM has higher prediction
accuracy than BPN by 6.2309%
 SVM performs better than CBR at 5% statistical
significance level
 SVM does not significantly outperform BP
 BP and CBR do not significantly outperform each
other
36