9. ARTIFICIAL NEURAL NETWORK
(ANN)
• Elegant & Powerful
• Approximate most complex
functions
• Blackbox
BIG DATA &
FAST PROCESSING
• Make nonlinear math solutions
feasible and imperative
• Realistic performance
BACKPROPAGATION
ALGORITHM
• Solve complex pattern at
its best
• Lower gradient to
minimize errors
10. RESEARCH
FRAMEWORK
Capital Market
Composite Stock Price
Index Historical Data (t)
Interest Rate
InflationInvestment
Macroeconomic Variables
Historical Data (t)
Data Mining
Prediction
Artificial Neural Network
Backpropagation
Composite Stock
Price Index (t+1)
Exchange Rate
Money Supply
12. PREPROCESSING
Outlier Replacement
Data Smoothing
Data Normalization
Data Partition
ANN DESIGN
PERFORMANCE
PARAMETER
1
2
3
4
ANN Architecture
ANN Parameter
Train, Validate, Test
Prediction
1
2
3
4
Accuracy (R-value)
Error (MSE)
1
2
13. Exchange Rate (ER)
Historical Data
Interest Rate (IR)
Historical Data
Inflation Rate (IF)
Historical Data
Money Supply (M2)
Historical Data
IDX Composite
Future Values
5
6
7
8
4
3
2
9
1
Bias Bias
Heaton (2008) in Kumar & Marugan (2017)
14. Dec ‘05 – Nov ‘16 Dec ’16 – Nov ‘17
Train (92), Test (20), and Validation (20) Phase Prediction Phase (12)
Data Sample (144)
Source: processed data result by researcher (2018)
15. NON PREPROCESSED PREPROCESSED
Under Fit Model
R-Value : 0,98842
Good Fit Model ✓
R-Value : 0,99826
----- Train
----- Validation
----- Test
----- Train
----- Validation
----- Test
16. Time Delay : 1 Time Delay : 2
Time Delay : 3 Time Delay : 6
Over Fit Model
Good Fit Model ✓
----- Train
----- Validation
----- Test
17. Time Delay : 1 Time Delay : 2 Time Delay : 3 Time Delay : 6
Good Fit Model ✓
R-Value : 0,99826
Good Fit Model ✓
R-Value : 0,99789
Over Fit Model
R-Value : 0,99972
Over Fit Model
R-Value : 0,99966
Model
(based on time delay)
Overall Accuracy
(R-value)
Overall Error
(MSE)
1 99,826% 0,002997392
2 99,789% 0,003634197
20. Conclusion
2. The best time to predict IDX Composite is two months before
the predicted month.
3. Macroeconomic variables are the proper indicators to predict
IDX Composite movements.
1. ANN predict the IDX Composite successfully
with final results of accuracy 96,38% and
0,0046 average errors.
21. Suggestion
1. Add or try another variables such as
microeconomic variables
2. Daily data is highly suggested since the
investors and issuers more require daily
prediction rather than monthly data. Daily
data also have more complex functions
which more challenging and beneficial.