TBB: data science training for bankers

62 hours

Duration

21 days, 7 modules

9+ / 10

Evaluation

TBB participant survey per module

10 / 10

Highest

Introduction and machine learning modules

2

Industry cases

Retail sales forecasting · bank transaction classification

Problem

Many people in banks work with data every day without using SQL or Python, and never see the models behind the reports. For the Banks Association of Türkiye’s training programme, our goal was to take a bank employee who uses data in their work from zero to the point of building a machine learning model themselves. So the programme was weighted towards practice: participants spent every evening at the keyboard.

Seven modules

We called the programme the “Data Analytics Journey”. Each module builds on the one before, and when the last one ends the participant has been through every step of a data science project.

#	Module	Hours	What was learned
1	Introduction to data analysis, data science and AI	2	Data literacy; what analysts and data scientists do; the tools used throughout the programme
2	Databases and SQL querying	12	Relational structures, data types; select, where, group by, join, subquery, DML; rank, case, while
3	Python programming	15	Data structures, conditions, loops, functions, files; pandas, numpy, matplotlib
4	Basic statistics and EDA in Python	3	Missing and outlying values, distributions, correlation; exploratory analysis
5	Data visualisation and tools	6	Visualisation principles; dashboards in Power BI
6	Basic machine learning in Python	18	Data preparation, regression and classification, validation, model selection
7	Data science industry cases	6	Two full cases, start to finish: retail store sales forecasting, bank customer transaction classification

Four instructors shared the modules: Ahmet Can taught SQL, Burak Python and machine learning, Hasan visualisation, Sabri the introduction and the industry cases. Sessions ran on weekday evenings from 17:00 to 20:00 over Teams. Each module’s homework was solved in the next session, code was shared in Google Colab and files were kept on a shared Drive.

Industry cases

The last module worked as a final exam: two real datasets, worked through from start to finish. In the retail case a store chain’s daily sales were forecast with calendar, promotion and competitor-distance variables; the work went from EDA all the way to model comparison. In the bank case a model was built to classify customer transactions from hundreds of anonymised variables, with the focus on class imbalance, feature selection and gradient boosting. Participants who had learned to write SQL in the first module were debating XGBoost parameters in the last.

Measurement

At the end of every module TBB ran a standard participant survey. It asked whether the training added to knowledge and skills, whether the exercises supported learning, and whether the presentation and materials were adequate. All seven modules scored above 8.5 out of 10; the introduction and machine learning modules got full marks, visualisation and EDA 9.6–9.9. The SQL module scored lowest. It was also the module with the densest content and the most homework, and the score led us to rethink SQL hours and homework load in later programmes.

Why this way

• Evening sessions were hands-on. Three hours after work is tiring; had the content been built as lectures, attendance would have dropped by the third week.
• The industry cases were known from day one. In the first session participants saw what they would be able to do by the end of the programme, so they knew why they were learning SQL.
• Survey scores were used to correct the programme. We decided which modules needed different hours and homework load by looking at the scores.

Related

• Kuveyt Türk: master–apprentice mentoring for the Digital Marketing and Retail Customer Analytics teams
• Defacto DSO: setting up a data science office and four years of consulting
• Kalekim: from MVP to a forecast management system in three years

All our work