HW 2 – MACHINE LEARNING FOR FINANCE
A.y. 2023/2024
 The assignment has to be presented in a document of up to ten pages, created using a word processor such as Microsoft
Word. The document may include tables, figures and similar items. The Working Group (WG) is required to format
the document as follows: a top margin of 2.5 cm and the other margins of 2 cm; Times New Roman font with size 12
pt; and single spacing. At the top left of the document, information relating to each student belonging to the WG must
be reported, including the surname and student number. The document has be created both in its original format, for
example docx, and in pdf. The necessary calculations has to be performed using a programming environment, such as
MATLAB.
 The document, the code(s), and the dataset(s) used have to be placed in a single folder. This folder must be
compressed, and the identifier for the compressed folder should be: ‘Last_name_of_student_1-
Last_name_of_student_2-HW2’ (e.g., ‘Corazza-Costola-HW2’).
 The compressed folder has be uploaded by a single WG student into the 'HOMEWORKS' sub-folder present in the
course folder in Google Drive no later than ... [To be decided together].
[1] The Working Group (WG) considers the MATLAB code BAG_RF_2023_2024.m and two of the time
series of daily closing stock prices downloaded for carrying out the first homework. Then, for each time
series the WG builds a dataset of input-output pairs in which: the inputs are the current and the four past
returns; the output is the sign of the one-step ahead return.
The WG searches for the setting of the parameters pe, num_trees, min_leaf and max_split which
provides a “good” classification of the sign of the one-step ahead returns. The WG presents the
development and comments the results.
[2] The WG considers the MATLAB code BOOSTING_2023_2024.m and the same two time series used
in the exercise above. Then, again as in the exercise above, for each time series the WG builds a dataset of
input-output pairs in which: the inputs are the current and the four past returns; the output is the sign of the
one-step ahead return.
The WG searches for the setting of the parameters pe, d_split, num_trees and shrinkage which
provides a “good” classification of the sign of the one-step ahead returns. The WG presents the
development, comments the results and compare them with those obtained in the exercise above.
[4] The WG considers the MATLAB code mlp_01_3.m and the same two time series used in Exercises 1
and 2.
The WG searches for the setting of the parameters n_input, “number of the hidden layers”, “number of
nodes per hidden layer”, net.divideParam.trainRatio, net.divideParam.valRatio,
net.trainParam.lr and of the training function in {trainlm, trainbfg} which provides a “good”
learning. The WG presents the carrying out, comments the results and compare them with those obtained in
Exercises 1 and 2.