Sherif Morsi
Open to 4–16 month internships and co-ops across Canada Canadian Permanent Resident

Sherif Morsi.

Third-year Computer Science student at Concordia University. I apply machine learning (ML) to economic research, market shocks and financial analysis, test every model against simple baselines on data it has never seen, and publish the results that disappoint next to the ones that don't.

See the work Email me GitHub LinkedIn
Selected work

Three research systems, measured honestly.

Each starts with a question an economist or strategist would ask, splits its data by date so nothing from the future leaks in, and reports where the models fell short.

Zero-shot forecast error on held-out datasets falling as model size grows, for three transformers pretrained from scratch
01Deep learning · Forecasting · Research engineering

Time-series foundation model, from scratch

Can a small transformer, pretrained from scratch on public data checked for overlap with the benchmark, forecast series it has never seen?

A decoder-only transformer, its tokenizers, training loop and sampling written by hand in PyTorch, pretrained on a free Colab GPU and scored zero-shot with the official GIFT-Eval code. An audit of the pretraining corpus found copies of benchmark series inside it; every one was removed before training. Evaluation rules were fixed in git before scoring.

0.605zero-shot CRPS relative to Seasonal Naive on held-out datasets, down from 0.740 for the smallest model
801pretraining series found to overlap the benchmark, removed before training
10.1Mparameters, every layer written by hand in PyTorch
What didn't work

Prediction intervals are too narrow: the 80% interval covers 65% of outcomes, because long forecasts feed back medians. The first GPU run also diverged in half precision until extreme values were clipped.

  • Python
  • PyTorch
  • GluonTS
  • statsforecast
  • NumPy
  • pandas
  • pytest
  • Hypothesis
  • GitHub Actions
Case study Code and results Research note ↗
World Shock Lab scenario composer: a Federal Reserve hike combined with an oil supply cut, with the response of Canadian bond yields, the Canadian dollar and Canadian stocks over 20 trading days
02Rates and macro research · Causal inference · Live forecasting

World Shock Lab

How do the world's markets, and Canada's in particular, react when the Federal Reserve (Fed) or the Bank of Canada (BoC) surprises them?

An open-source research platform measuring how Fed, BoC and Organization of the Petroleum Exporting Countries (OPEC) surprises move 53 global assets. Estimators are validated by Monte Carlo simulation and by replicating published papers; a machine learning benchmark was pre-registered in git; a scenario tool and forecasts published before each central-bank decision sit on top.

−1.10%S&P/TSX Composite on the day of a 10 basis-point Fed surprise, vs −0.19% for the BoC
11 of 13published findings replicated (sign and 90% band)
0 of 80pre-registered regime tests survive false-discovery-rate control
What didn't work

Machine learning did not beat a linear model on 168 held-out announcements: error difference −0.003, 90% interval −0.019 to +0.014.

  • Python
  • PyTorch
  • statsmodels
  • scikit-learn
  • DuckDB
  • React
  • TypeScript
  • D3.js
  • GitHub Actions
Live site Case study Code ↗
ATLAS world atlas: a map of real gross domestic product growth by country, with Japan's economic indicators in a side panel
03Economic systems · Machine learning · Full-stack engineering

ATLAS

How does a cost shock in one industry spread through the global economy?

A local-first economic research terminal. It connects 42,900 World Bank observations for 177 countries and territories with seven Organisation for Economic Co-operation and Development (OECD) input-output tables, solves combined production-cost shocks on a 405-node network with PyTorch, and replays 12 trained PyTorch models against held-out data, all from one FastAPI process on your own machine.

405production-network nodes across 81 areas, solved and differentiated with PyTorch
12trained PyTorch checkpoints replayed by the release gate
44 + 27backend and browser tests passed for release v0.1.2
What didn't work

The neural models did not beat their baselines: ridge regression was best in the 2022 graph experiment, and the recurrent model did not consistently beat ridge on 10 held-out years of US growth.

  • Python
  • PyTorch
  • FastAPI
  • DuckDB
  • CVXPY
  • React
  • TypeScript
  • MapLibre
Case study Download release Code ↗
How I work

Evidence first.

The same four habits run through both projects. They are why the numbers above can be trusted, including the disappointing ones.

Time-based holdouts

Nothing from the future leaks in.

Data are split by date, never at random. World Shock Lab trains up to 2014, validates on 2015–2018 and tests on 2019–2026; ATLAS's graph models train on 2018–2020, validate on 2021 and test on 2022.

Baselines

Every model faces a simple rival.

A neural network against a linear projection and a forecast of no move. Recurrent and graph models against ridge regression, the training mean and persistence.

Pre-registration

The test is fixed before the result.

In World Shock Lab the splits, metric and decision rule were committed and tagged in git before any model was fitted, the held-out test set was scored once, and live forecasts are published before each decision and never edited.

Honest reporting

Null results are results.

World Shock Lab's machine learning test found no gain over a linear model, and ATLAS's neural models did not beat ridge regression. Both say so up front, with the numbers.

Experience

Industry experience.

Machine learning for telecommunications network energy, and credit analysis for small and medium-sized businesses.

Aug 2026 – Sep 2026Egypt (hybrid)

Artificial Intelligence (AI) & Digital Technologies Intern

Orange Egypt × Nokia Egypt · Hybrid Internship Program

  • Built and trained machine learning models in Python and PyTorch, applying predictions to energy-optimization actions in a team-based fifth-generation (5G) network project.
  • Analyzed high-velocity network telemetry for predictive modeling and to identify operational indicators of energy usage and network efficiency.
  • Applied AI-driven automation to turn model predictions into data-driven energy-optimization actions.
  • Led my team in presenting the project proposal, and presented model results and technical findings while collaborating across AI, data and telecommunications.

Trained through sessions led by industry professionals, workshops and mentorship in mobile networks from the second to the fifth generation (2G to 5G), radio network planning, AI and machine learning, data engineering and cybersecurity.

Aug 2025 – Sep 2025Cairo, Egypt

Credit Analyst Intern, Small and Medium-sized Enterprise (SME) Credit

Qatar National Bank (QNB)

  • Shadowed and worked with the SME credit analyst team across the full credit process, mainly on agriculture and technology companies.
  • Studied credit proposals and the business plans behind them, and shadowed the team's initial review of credit facility proposals submitted by SME Banking units for approval.
  • Shadowed and supported creditworthiness assessments of counterparties requiring credit limits for operational needs.
  • Spread SME clients' financial statements under supervision, drafted sections of credit memos for review by the credit team, and researched clients and counterparties to support credit assessments.
Activities

Beyond the classroom.

A running community I cofounded, trading competitions and a finance book club.

Sep 2024 – Present · Montreal

Cofounder, Arab Run Club Montreal

4,500+ followers

Cofounded the club and grew it to over 4,500 followers, organizing weekly runs in Montreal.

Sep 2026 – Present · Concordia University

Member, Quantitative Research and Competitions Club (QUARCC)

Competing in campus rounds to represent Concordia at the Global Market Making Games, an international market-making competition hosted by the Amsterdam Investment Club that tests mental maths, probability, pricing, risk-taking and market making. Trades in the club's trading competitions and attends its game nights.

Sep 2026 – Present · Concordia University

Member, John Molson Book Club

Reads one finance-related book each month with the club, joins its weekly discussions and workshops, and recommends books.

About

Making financial problems testable.

I'm a third-year Computer Science student at Concordia University in Montreal, enrolled in its co-op program. I'm interested in applying AI and machine learning to economic research, macro and market shocks, and financial analysis.

My projects follow the same path end to end: a data pipeline with provenance, models compared against simple baselines on data they never saw, and an interface people can use to explore the results.

Based in
Montreal, Quebec, Canada
Status
Canadian Permanent Resident · United Arab Emirates (UAE) Golden Visa holder
Available
4, 8, 12 or 16-month internships and co-ops in any Canadian city; willing to relocate
Languages
English (fluent) · Arabic (fluent) · French (learning)
Interests
Financial markets, investing, finance books, running
Education

Concordia University · Bachelor of Computer Science

Gina Cody School of Engineering and Computer Science · Co-op program · Expected May 2028

Relevant coursework: Machine Learning, Artificial Intelligence, Financial Markets, Data Structures and Algorithms, Linear Algebra, Multivariable Calculus.

Programming

Python, Java, C++, JavaScript, TypeScript, Structured Query Language (SQL)

Machine learning

PyTorch, TensorFlow, scikit-learn, Medical Open Network for Artificial Intelligence (MONAI), machine learning, deep learning

Data

pandas, NumPy, SciPy, DuckDB, SQLite, Parquet, data analysis, predictive modeling, data preprocessing

Generative AI

LangChain, prompt engineering, retrieval-augmented generation (RAG), some hands-on use

Finance

Excel financial modeling and valuation, PowerPoint; from a supervised credit internship: SME credit assessment, credit proposal review, counterparty creditworthiness assessment, financial statement spreading, credit memo drafting

Optimization

CVXPY, linear programming

Software engineering

Object-oriented programming, algorithms, FastAPI, Representational State Transfer (REST) Application Programming Interfaces (APIs), React, Vite, Next.js, D3.js, Git, GitHub, pytest, Playwright, unit testing, Continuous Integration and Continuous Delivery (CI/CD) with GitHub Actions, Docker, Figma

Spreadsheets and Business Intelligence (BI)

Excel formulas, lookup functions, PivotTables, charts, data cleaning, macros in Visual Basic for Applications (VBA), Power Query, Power BI

Networks

5G networks, telecommunications, network telemetry, automation

Contact

Let's build something
measurable.

I'm looking for internships and co-ops of 4 to 16 months anywhere in Canada, especially where machine learning meets economics and finance. Email is the fastest way to reach me.

LinkedIn GitHub