QBUS2820 Predictive Analytics
Individual Assignment 1
Key information
1. Required submissions (through Canvas/Assignments/Individual Assignment 1)
a. ONE written report (word or pdf format)
b. ONE Jupyter Notebook .ipynb
Please upload both files to canvas in the SAME submission, as separate files (NO
zip file).
2. Due date/time and closing date/time: See Canvas. The late penalty for the
assignment is 5% of the assigned mark per day, starting after 23.59pm on the due
date.
3. Weight: 30% of the total mark of the unit.
4. Length: The main text of your report should have a maximum of 10 pages with the
usual font size 11-12. You should write a complete report including sections such as
business context, problem formulation, data processing, Exploratory Data Analysis
(EDA), methodology, analysis, conclusions and limitations, etc.
5. If you wish to include additional material, you can do so by creating an appendix.
There is no page limit for the appendix. Keep in mind that making good use of your
audience’s time is an essential business skill. Every sentence, table and figure have
to count. Extraneous and/or wrong material will reduce your mark no matter the
quality of the assignment.
6. Anonymous marking: As the anonymous marking policy of the University, please
only include your student ID in the submitted report, and do NOT include your
name. The file name of your report and code file should follow the following format.
Replace "SID" with your Student ID. Example: SID_Qbus2820_Assignment1.
7. Presentation/clarity is part of the assignment. Markers will allocate 10% marks for
clarity of writing and presentation. Numbers with decimals should be reported to
the fourth decimal point.
Key rules:
Carefully read the requirements for each part of the assignment.
Please follow any further instructions announced on Canvas.
You must use Python for the assignment. Use "random_state= 1" when needed, e.g.
when using “train_test_split” function of Python. For all other parameters that are not
specified in the questions, use the default values of the corresponding Python
functions.
Reproducibility is fundamental in data analysis, so that you will be required to submit a
Jupyter Notebook that generates your results. Not submitting your code will lead to a
loss of 50% of the assignment marks.
2
The notebook must run without errors and produce results consistent with the report
when accessed through Kernel -> Restart & Run All from the Jupyter menu, assuming
that the train and test datasets are in the same folder as the notebook. Failure to do so
can results in a loss of up to 50% of the assignment marks.
Failure to read information and follow instructions may lead to a loss of marks.
Furthermore, note that it is your responsibility to be informed of the University of
Sydney and Business School rules and guidelines, and follow them.
The Task
You will work on a Credit Risk Rating Data set. This is a dataset about credit ratings given to
a list of publicly traded firms in the US, gathered from 2014 to 2015. The dataset consists of
multiple financial variables of the firms, and their respective rating given to the firms by the
rating agency Standard and Poor’s.
The assignment consists of applying models and model selection methodologies to arrive at
models that predict the rating from some of the other variables measured.
The credit ratings are often given in an ordered scale, from AAA to D, but in our dataset, the
ratings have been grouped and transformed to numbers, integers from 1 (the group of best
rankings) to 4 (the group of worse rankings).
The dataset (`credit_data.csv`) comes from a research paper that explores performance of
‘Artificial Intelligence’ methods for predicting credit ratings:
https://doi.org/10.1016/j.eswa.2020.113925
You might read the introduction of the paper for motivation and context. In addition to the
original variables which should be self-explanatory, the following variables have been
added:
`Rating`: The credit risk rating transformed to numbers
`ID`: Unique ID identifying the firm
`Year`: Year the report was made (they are all done on Q4 of that year).
1. Problem description
A primary goal is finding a model that is accurate in predicting the rating of the firms. The
accuracy of the predictions is initially measured in Mean Absolute Error (MAE).
A secondary goal is to get an understanding of which are the main factors that drive the
ratings, according to the model, this would require that at least one of the models uses a
few variables or that you can create a coherent explanation out of one of the models if all
use many variables (you do not need to be a finance expert for this, though if you want to .
Select three models, one from each model family to predict the target variable Rating.
These model families are:
a linear regression model,
a kNN regression model,
A third model. This model can be any model of your choice that is not linear
regression nor kNN (might even be a model not covered in the QBUS2820 unit). This
is to encourage you to self-explore and self-study, since the ability of self-study is
critical in the field of machine learning which is evolving rapidly.
All the models need to be fine-tuned with hyperparameter search (when appropriate) and
potentially variable selection. The methodology should maximize the predictive accuracy
and first, and the explanation second. When the three models have been tuned, you will
compute an accurate estimate of the prediction error of these models and make a final
decision among the three. In the conclusions, you also have to explain the driving factors of
the ratings, if the chosen model is not explainable, then use another (or several) and
carefully justify the tradeoffs (accuracy sacrificed vs explanations) .
The model selection part of the assignment, including:
intro/business context/problem formulation
exploratory data analysis
The three models
The conclusions section
Represents the main body of the report and makes 80% of the grade of the assignment.
In addition to the model selection above, the following short exercises. Create a section for
each of the questions and remember to explain and discuss the methodology in the report
as weel as in the main body.
(5%) Find the best predictive model that uses a single predictor (only one variable),
you can use all model classes .
(10%) Think a bit more carefully about the implications of the error function used in
the main part of the assignment, the Mean Absolute Error, and the interpretation of
the response variable ‘Ratings’. Describe a more ‘appropriate’ objective function
that considers the differences between predictions and true values, and the
implications of these differences. Program this function in python. Sketch it using a
table or plot. Re-evaluate your candidate models according to this new function and
comment on the difference in results (if any). You do not need to re-train your
models for this new error function (in practice we would try to).
? (5%) Notice the ID and year variable. What is the main problem that these
represent, with respect to the basic assumptions we make in the predictive analytics
setting (the main violation of the assumptions required to do predictive analytics)?
How could you transform the dataset to solve or mitigate the effects of this
problem?
The grading of the assignment will be based on the methodology and justifications,
removing points for methodological errors, incomplete sections, etc. There is no ‘minimum’
predictive accuracy to be reached, but you need to apply a good methodology.
2. Written report
The purpose of the report is to describe, explain, and justify your solution. Be concise and
objective. Find ways to say more with less. When in doubt, put it in the appendix. Below are
some guidelines on how to work on the Task.
Preparation. You read and understood the assignment requirements and are aware that
this is part of the assessment. You understand that machine learning is grounded in
rigorous logic and theory that should inform your practical analysis. You understand that
there is no single right solution and that trying different approaches and discovering
empirically what works best for a particular problem is natural and desirable in this type of
analysis.
Business context and problem formulation. The report includes a discussion of the context
for the analysis, the problem and questions/hypotheses to be addressed, and how you plan
to measure the success of your proposed solutions.
Data processing. You make sure that the dataset is free of errors and correctly processed
for your analysis. You handle missing values and other issues appropriately. You describe
the data processing steps in a clear and concise way.
Exploratory data analysis (EDA). Your report describes your EDA process, presenting only
selected results. You studied key variables individually. You note any features of the data
that are relevant for model building (some variables might be ‘invalid’ for predictive
purposes). You note the presence of outliers and any other anomalies that can affect the
analysis. You explain the relevance of the EDA results to your subsequent modelling. Your
EDA section in the report is concise, leaving additional figures and tables to the appendix if
needed. Outliers should be clear (e.g. negative values for counting variables). EDA is not the
place to do variable selection and outliers of a non clear nature (e.g. very large values)
should be either not removed or further analyzed using the predictive model performance.
The dataset has many variables and you are not expected to report on all of them
individually, just report your methodology and main findings.
Variable selection. You describe and explain your process for variable selection. Your
choices are justified by data analysis and/or trial and error. Other than potentially invalid
variables from the dataset, the decision should be driven by the performance of the models,
not based on opinions (you are free to comment on the disagreements between your
background knowledge and the models).
Methodology and modelling. You clearly describe and justify the models, methods, and
algorithms in your analysis. The choice of methods is logically related to the assignment
requirements, the substantive problem, underlying theoretical knowledge, and data
analysis. This may involve systematic trial and error, but the report should focus on your
final solutions. Your methodology pays attention to statistical variability. You report all
crucial assumptions and check them as relevant via formal and informal diagnostics. You
clearly recognize when an assumption is not satisfied or questionable. Some problems may
be unfixable given the available data and methods. In this case you can identify what
additional information or methodology could allow you to fix these problems.
Analysis and conclusions. Your analysis is rich. You correctly interpret the results and
discuss how they address the substantive question. The reasoning from methodology and
results to your conclusions is logical and convincing. You are not misled by overfitting. Your
analysis pays attention to statistical variability. You make no claims for which you have no
evidence. You do not make statements that imply causation when discussing associations.
You explicitly acknowledge when limitations of the data or methods lead to uncertainty
about your answer to the substantive question.
Writing. Your writing is concise, clear, precise, and free of grammatical and spelling errors.
You use appropriate technical terminology. Your paragraphs and sentences follow a clear
logic and are well connected. There is a clear distinction between the essential parts of the
report and less important material (use the appendix). Your text refers to meaningful names
for variables and subjects. If you use an abbreviation or label, you first have to define it.
Report. Your report is well organized and professionally presented and formatted, as if it
had been prepared for a client later in your career. There are clear divisions between
sections and paragraphs.
Tables. Your tables are appropriately formatted and have a clear layout. The tables have
informative rows and column labels. The tables are as much as possible easy to be
understood on their own (in the real world, a significant part of your audience will skim-read
by going straight to the tables). The tables do not contain information which is irrelevant to
the discussion in your report. Your table is not an image. The tables are placed near the
relevant discussion in your report. There is no text around your tables.
Figures. Your figures are easy to understand and have informative titles, captions, labels,
and legends. The figures are well formatted and laid out. The figures are placed near the
relevant discussion in your report and are references from the text of the report. Your
figures have appropriate definition and were directly saved from Python into an image file
format. There is no text around your figures.
Numbers. All numerical results are reported to four-decimal point.
Referencing. You add citations for your sources. The references follow a recognizable style
(e.g. the Harvard Referencing System, MLA, APA, Vancouver, etc.)
Python code. The code is presented in a neat and compact way. The code uses meaningful
variable names and can be easily followed by someone with training in Python and statistics.
Someone should be able to run your code and reproduce all the results that appear in your
report. Your code has comments that clearly indicate which parts correspond to which
sections of your report. You explicitly acknowledge when you borrow pieces of code from
sources other than the lecture and tutorial materials.