Skip to content
← Back to projects

Projects · Forage job simulations

Virtual internships

Twelve Forage job simulations · Oct 2022 – Aug 2023

What twelve company-designed job simulations added to how I work with data, and what I actually produced.

Self-paced · in my own time

See all twelve ↓Back to projects

These were not jobs. Each Forage job simulation is a free set of tasks that a company publishes for students, and I worked through twelve of them in my own time. Only four of my repos keep my own work: KPMG, British Airways, Tata and PwC Digital Intelligence. The other eight keep only the briefs because I submitted the work on Forage, so for those I describe the tasks as set and claim no results. The repos are private, and no logos, company data, briefs or certificates are hosted here.

Programmes
12
Companies
11
Period
Oct 2022 – Aug 2023
My work kept
4 of 12 repos

What the set adds up to

Twelve short simulations do not make anyone a consultant or a data scientist, and I don't claim they did. What they gave me was repeated practice at the parts of the job that coursework rarely asks for: a business question owned by a real decision maker, data that has to be checked before it can be trusted, and an answer that has to fit on one slide, in one email or in a short video.

Across eleven companies the same working rhythm kept coming back, and it is still the order I work in today.

  1. Frame the question first

    Most briefs open with a stakeholder and a decision, not a data set. For Tata I wrote down eight questions for the CEO and CMO before drawing a chart, and for KPMG I scoped the work in three phases before any modelling.

  2. Check the data before trusting it

    For KPMG my first deliverable was an email listing nine data quality issues. For Tata I removed returns and invalid prices before building any view, and the GE brief hinges on one airline that stored a temperature on a different scale.

  3. Model, then explain

    For British Airways I went from scraped reviews to a booking model, and for PwC I trained a model of which bank clients to call and used SHAP to explain it, for the whole test set and for single clients. BCG, Cognizant and Standard Bank practise churn, stock and credit risk models as set.

  4. End on one page

    Almost every programme ends in a single slide, a client email or a short recorded presentation for someone without a technical background. Writing that page is the skill I use most from the set.

Skill map

Each row is a programme and each column is a skill area it asked for. A filled square means my own work for that area is kept in the repo. A hollow square means the brief asked for it but only the brief is kept, so I claim nothing beyond the task.

Skill areas covered by each of the twelve Forage programmes
ProgrammeFramingPrepAnalysisModelsVisualsComms
KPMGData AnalyticsBusiness framing: My work, kept in the repoData preparation: My work, kept in the repoAnalysis and experiments: In the brief, work not keptModelling: In the brief, work not keptVisualisation: In the brief, work not keptStakeholder communication: My work, kept in the repo
GE AviationData AnalyticsBusiness framing: Not part of the programmeData preparation: In the brief, work not keptAnalysis and experiments: Not part of the programmeModelling: Not part of the programmeVisualisation: In the brief, work not keptStakeholder communication: Not part of the programme
TataData VisualisationBusiness framing: My work, kept in the repoData preparation: My work, kept in the repoAnalysis and experiments: My work, kept in the repoModelling: Not part of the programmeVisualisation: My work, kept in the repoStakeholder communication: In the brief, work not kept
Accenture North AmericaData Analytics and VisualisationBusiness framing: In the brief, work not keptData preparation: In the brief, work not keptAnalysis and experiments: In the brief, work not keptModelling: Not part of the programmeVisualisation: In the brief, work not keptStakeholder communication: In the brief, work not kept
PwC SwitzerlandPower BIBusiness framing: In the brief, work not keptData preparation: Not part of the programmeAnalysis and experiments: In the brief, work not keptModelling: Not part of the programmeVisualisation: In the brief, work not keptStakeholder communication: In the brief, work not kept
QuantiumData AnalyticsBusiness framing: Not part of the programmeData preparation: In the brief, work not keptAnalysis and experiments: In the brief, work not keptModelling: Not part of the programmeVisualisation: In the brief, work not keptStakeholder communication: In the brief, work not kept
Red BullOn-Premise SalesBusiness framing: Not part of the programmeData preparation: Not part of the programmeAnalysis and experiments: In the brief, work not keptModelling: Not part of the programmeVisualisation: In the brief, work not keptStakeholder communication: In the brief, work not kept
British AirwaysData ScienceBusiness framing: Not part of the programmeData preparation: My work, kept in the repoAnalysis and experiments: My work, kept in the repoModelling: My work, kept in the repoVisualisation: My work, kept in the repoStakeholder communication: My work, kept in the repo
PwCDigital IntelligenceBusiness framing: Not part of the programmeData preparation: My work, kept in the repoAnalysis and experiments: Not part of the programmeModelling: My work, kept in the repoVisualisation: My work, kept in the repoStakeholder communication: Not part of the programme
BCGData Science & AnalyticsBusiness framing: In the brief, work not keptData preparation: In the brief, work not keptAnalysis and experiments: In the brief, work not keptModelling: In the brief, work not keptVisualisation: Not part of the programmeStakeholder communication: In the brief, work not kept
CognizantArtificial IntelligenceBusiness framing: In the brief, work not keptData preparation: In the brief, work not keptAnalysis and experiments: In the brief, work not keptModelling: In the brief, work not keptVisualisation: Not part of the programmeStakeholder communication: In the brief, work not kept
Standard BankData ScienceBusiness framing: In the brief, work not keptData preparation: In the brief, work not keptAnalysis and experiments: In the brief, work not keptModelling: In the brief, work not keptVisualisation: Not part of the programmeStakeholder communication: In the brief, work not kept
Programmes7/1210/1210/126/129/1210/12
  • My work, kept in the repo
  • In the brief, work not kept
  • Not part of the programme

The twelve, by skill area

Each card gives the business question, what I delivered or what the tasks asked for, and the skills it built. Dates come from my own commits in each repo.

Data quality and joins

Two programmes made the data itself the deliverable: finding what is wrong with it, and joining sources that do not agree.

KPMG · Data Analytics

Oct – Nov 2022

My work kept

Business question
Sprocket Central, a bike and cycling accessories retailer, shared three tables of customer and transaction data and a list of 1,000 new customers. Which of the new customers should its marketing team target first?
What I delivered
I profiled the four worksheets in pandas and wrote the client an email that set out nine data quality issues, such as missing values, inconsistent gender entries, unnamed columns and a date of birth in 1843. For the second module I drafted a short deck that split the work into exploration, model development and interpretation, suggested external data such as ABS population figures, and proposed ranking metrics to judge the recommendations. The third module, a client dashboard, is not kept in the repo, so I claim nothing for it.
What I produced
  • Data quality email
  • Exploration notebook
  • Approach deck
Skills it built
Treating data quality as the first deliverable, and writing it up for a client as specific issues with what each one means for the analysis.
Tools
  • Python
  • pandas
  • Jupyter
  • PowerPoint

GE Aviation · Data Analytics

Nov – Dec 2022

Brief only · no results claimed

Business question
How do you combine engine, manufacturing and airport data into one table analysts can use, and are machined parts staying within their design tolerances?
Tasks as set
The first module asks for the flight engine data of four airlines to be merged into one table, after adjusting a temperature column that one airline stored on a different scale, with a bonus step that joins the supply chain and bill of materials sheets with lookups. The second asks for a Tableau run chart and KPI tables that show whether each operation's measurements fall inside specification.
Deliverables the brief asked for
  • Merged data set
  • Run chart
  • KPI tables
Skills the tasks practise
Spotting a unit mismatch before a merge, and reading a run chart to judge a process over time.
Tools
  • Excel
  • Tableau

Visualisation and storytelling

Three programmes were about choosing the right view for a decision maker and telling the story around it.

Tata · Data Visualisation

Oct 2022

My work kept

Business question
The CEO and CMO of an online retailer wanted to know what drives revenue and where the business should expand next.
What I delivered
Before drawing anything, I wrote down eight questions the two leaders were likely to ask, four quantitative and four qualitative, and explored the public UCI Online Retail data to see which ones it could answer. I then chose a chart type for each of five scenarios and wrote down why, such as a line chart for monthly seasonality and a filled map for demand by country. Finally I cleaned the data in pandas by removing returns and invalid prices, which took it from about 541,900 rows to 530,100, and built the four requested views as a Tableau workbook with a dashboard. The last module, a recorded presentation, is not kept in the repo.
What I produced
  • Question list
  • Chart choices with reasons
  • Cleaning notebook
  • Tableau workbook
Skills it built
Writing the decision maker's questions first and choosing each chart for the question it answers, rather than starting from the data.
Tools
  • Python
  • pandas
  • Tableau

Accenture North America · Data Analytics and Visualisation

Nov 2022

Brief only · no results claimed

Business question
A client brief sets a business problem about content and how users react to it. Which of seven related data sets answer it, and how do you present the answer to the client?
Tasks as set
Four modules. The first is a quiz on the business problem, the requirements and who on the team does what. The second asks for the relevant tables to be cleaned and merged on their keys into one final data set. The third asks for visualisations and a slide deck that tells a clear story, and the fourth for a recorded presentation to the client.
Deliverables the brief asked for
  • Merged data set
  • Slide deck
  • Video presentation
Skills the tasks practise
Reading a client brief for its requirements first, then cutting a data model down to the tables that answer them.
Tools
  • Excel
  • PowerPoint

PwC Switzerland · Power BI

Nov 2022

Brief only · no results claimed

Business question
A telecom client wants clear answers on three things: how its call centre is performing, which customers are at risk of leaving, and why gender balance in its executive team is not improving.
Tasks as set
After an introduction, three case tasks each ask for KPIs worked back from a stakeholder's question and a Power BI dashboard that shows them. The call centre task covers satisfaction, answered and abandoned calls and speed of answer. The retention task adds a short email of findings and suggested changes, and the diversity task asks for hiring, promotion, performance and turnover measures and the likely root causes of slow progress.
Deliverables the brief asked for
  • Three dashboards
  • Retention email
  • Root cause notes
Skills the tasks practise
Working back from a stakeholder's question to the KPI that answers it, and only then to the visual.
Tools
  • Power BI

Commercial analytics

Two programmes tied the analysis to a commercial call: whether a store trial worked, and how to win back an account.

Quantium · Data Analytics

Oct – Nov 2022

Brief only · no results claimed

Business question
Who buys chips at a large supermarket chain, what drives their spending, and did a new layout trialled in three stores lift sales?
Tasks as set
Three modules. The first asks for checks and cleaning of transaction and customer data, derived features such as pack size and brand, and customer segments that lead to a recommendation. The second asks for a control store for each trial store, chosen with a measure such as correlation or magnitude distance, and a test of whether sales changed during the trial and what drove it. The third asks for a client report built on the Pyramid Principle.
Deliverables the brief asked for
  • Customer segments
  • Trial and control analysis
  • Client report
Skills the tasks practise
Setting up a trial and control comparison, and leading a client report with the answer rather than the method.
Tools
  • Python
  • R

Red Bull · On-Premise Sales

Nov 2022

Brief only · no results claimed

Business question
Which bar and restaurant accounts are growing or slipping, and how do you win back an account whose sales have dropped?
Tasks as set
The first module asks for the compound annual growth rate of sales volume over five years of randomly generated account data in Excel, comparisons by account type, year, promotion programme and product range, and a short deck of findings. The second asks for a short recorded video message that answers a bar owner's objections on price and demand, using an active listening framework.
Deliverables the brief asked for
  • CAGR analysis
  • Findings deck
  • Video message
Skills the tasks practise
Reading account-level trends with a commercial eye, and answering an objection by listening and reframing before making a case.
Tools
  • Excel
  • PowerPoint

Models and explainability

Five programmes asked for a predictive model. The two I kept go from raw data to an explained result, and the other three practise churn, stock and credit risk models as set.

British Airways · Data Science

Nov 2022

My work kept

Business question
What do travellers say about the airline in public reviews, and which factors make a customer go on to complete a booking?
What I delivered
I wrote a threaded scraper with requests and BeautifulSoup that walked the Skytrax A to Z of airline reviews and saved reviews for 485 airlines, about 89% of those listed. I cleaned the 1,200 British Airways reviews, labelled their sentiment with a pretrained Hugging Face model, and summarised aircraft, traveller type, seat class and sentiment on one slide. For the booking task I prepared 50,000 booking records, fitted an OLS baseline and a neural network classifier, and checked how each feature related to a completed booking. That summary also went on one slide.
What I produced
  • Review scraper
  • Sentiment notebook
  • Booking model notebook
  • Two one-slide summaries
Skills it built
Collecting and cleaning messy web text myself, and fitting each analysis onto one slide for a non-technical reader.
Tools
  • Python
  • requests
  • BeautifulSoup
  • pandas
  • Hugging Face Transformers
  • scikit-learn
  • statsmodels
  • Plotly

PwC · Digital Intelligence

Aug 2023

My work kept

Business question
A large bank uses a model to decide which clients to phone about term deposits. Management wants to know how such a model reaches its decisions, overall and for individual clients.
What I delivered
I trained a cross-validated logistic regression on the public UCI Bank Marketing data, 41,188 campaign records, and checked it with a confusion matrix and a classification report. I then used SHAP to rank the features that drive predictions across the test set, and drew force plots that explain the predictions for the two clients the brief named. Economic indicators, such as the number employed and the three-month Euribor rate, led both explanations. The programme lists further modules, but the repo holds only this responsible AI task, so that is all I claim.
What I produced
  • Classification model
  • Global SHAP summary
  • Two local explanations
Skills it built
Treating an explanation, for the whole model and for one person, as part of finishing a model rather than an extra.
Tools
  • Python
  • pandas
  • scikit-learn
  • SHAP

BCG · Data Science & Analytics

Nov 2022

Brief only · no results claimed

Business question
PowerCo, a European gas and electricity utility, is losing small and medium business customers. Is price sensitivity driving that churn, and would a 20% discount keep the customers most at risk?
Tasks as set
Four modules take the question from a hypothesis to a steering committee. They ask for an email to the associate director on the data and models needed to test it, an exploratory analysis of customer and price data with a definition of price sensitivity, features built on the gap between December and January off-peak prices, a random forest that predicts churn with a justified choice of metrics, and a one-slide executive summary.
Deliverables the brief asked for
  • Email to the associate director
  • EDA notebook
  • Random forest model
  • Executive summary slide
Skills the tasks practise
Turning a commercial hypothesis into a testable data science problem, and judging a model by what it means for the client's bottom line.
Tools
  • Python
  • pandas
  • scikit-learn

Cognizant · Artificial Intelligence

Nov 2022

Brief only · no results claimed

Business question
Gala Groceries sells highly perishable stock. Can sales and sensor data predict hourly stock levels well enough to guide what it orders from suppliers?
Tasks as set
Five modules run from exploration to production. They ask for an exploratory analysis of sample sales data summarised in an email, a one-slide plan for which tables to use from the data model, and a model that combines sales, stock and temperature data, explained on one slide in business terms. They then ask for a documented Python module that trains the model and reports its performance, and finish with a quality assurance quiz on improving the deployed model.
Deliverables the brief asked for
  • Email
  • Planning slide
  • Model and results slide
  • Python module
Skills the tasks practise
Moving a model out of a notebook into a documented module another team can run, and describing its performance without technical metrics.
Tools
  • Python
  • pandas
  • scikit-learn
  • Google Colab

Standard Bank · Data Science

Nov 2022

Brief only · no results claimed

Business question
Loan officers take two to three days to process each home loan application. Can machine learning predict whether an applicant will default, so they get an answer as soon as they apply?
Tasks as set
The modules follow CRISP-DM. They ask for a multiple-choice SQL quiz on the data understanding stage, a notebook that answers the home loans manager's questions about the data and compares AutoML with a bespoke model, a presentation built around the data science life cycle, and a five to ten minute video for a manager with a limited technical background.
Deliverables the brief asked for
  • SQL quiz
  • AutoML vs bespoke notebook
  • Presentation
  • Video
Skills the tasks practise
Weighing AutoML against a hand-built model, and presenting the trade-off to a sponsor who is not technical.
Tools
  • SQL
  • Python
  • AutoML

Looking back

Re-reading my own notebooks for this page, three things stand out that I would do differently now.

  • Check the baseline first

    Most bookings in the British Airways data were never completed, so my classifier's accuracy was close to what always guessing “no booking” would score. I now put the class balance and a naive baseline next to any score I report.

  • Leave out what you only learn afterwards

    My PwC model kept call duration as a feature, and the UCI notes warn that duration is only known once the call is over. For a real calling list I would drop it and explain the model again.

  • Answer the question that was asked

    For KPMG I proposed a recommender system, but the brief asked which of 1,000 new customers to target, and they had no purchase history. A ranked list of customers by likely value would have answered it more directly.

Company and programme names appear as text only. The repos are private, so there are no source links, and no logos, company-provided data, briefs, certificates or scores are hosted here.

Back to projects