What I want to build and understand
I grew up in Iraq, and I have always wanted to understand its economy better. Iraq has valuable oil resources, but having those resources has not made development straightforward. I want to look at the numbers, understand what they can tell me, and build a Python application that helps me explore the questions properly.
My approved thesis title is Iraq’s Economic Outlook: Building an Economic Analysis System with Iraq as a Case Study. I confirmed that the full title is approved on 13 October 2026. I am studying Economic Informatics at ASE Bucharest, and I plan to defend in July 2027. The final econometric specification and detailed scope remain for discussion with my professors.
Professor Uscatu has agreed to coordinate my thesis. I will discuss the final question and scope with him. Iraq gives the study its personal purpose, while the analysis system is something I want to develop further and eventually adapt to other countries and suitable datasets.
The questions behind the project
How has Iraq's economy changed since 2003? How much does the government rely on oil-related income? What happens when revenue falls but spending does not fall with it? Can monthly data help me make useful short-term forecasts?
I also want to understand why Iraq's development has followed a different path from countries such as the UAE and Saudi Arabia. Where did public spending go, and how much helped build a more diverse economy? These are questions I want to investigate. A few charts or a country ranking cannot answer them on their own.
I do not want to begin with a conclusion and then find numbers to support it. For example, Iraq's budget has not been in deficit every year. I need to check what happened, when it happened, and what the available evidence actually supports.
Why I start with Iraq after 2003
I chose 2003 because it marks the US-led invasion and the fall of Saddam Hussein's regime. I want to understand the economy in the period that followed, without assuming that one event explains every later outcome.
I use two parts of the data together. The yearly data gives me the longer development picture from 2003. The monthly government accounts let me look more closely at revenue, spending and short-term changes.
The yearly figures for growth, real GDP per person and manufacturing run through 2025. Real GDP per person is a measure of economic output per person, adjusted for price changes. It is not the same as the income of a typical household. The oil-rent figures are available through 2021. They estimate the value of oil production after production costs; they are not government oil revenue.
My consistent monthly forecasting dataset runs from January 2016 to June 2026: 126 months. I also have an earlier archive starting in 2008, but some years have changes in coverage or accounting adjustments that need more checking. That is why I am keeping the reliable modelling window separate for now.
Both frequencies can belong in the same thesis. I keep yearly figures yearly and monthly figures monthly. I do not turn one annual GDP value into twelve observations. I also keep January–June 2026 separate from a complete year.
How I prepare the monthly figures
The Central Bank of Iraq tables report running totals within each year. A February value includes January and February together. To get February's own amount, I subtract January's total. For March, I subtract February's total, and so on. January starts a new year, so it uses its own reported amount.
My code checks that these monthly amounts add back to the yearly total and that revenue minus spending equals the balance. It keeps the original values, source links, units and warnings so that I can trace a result back to its source.
Some differences can reflect a correction to the accounts. I keep those visible. For example, one oil-and-mineral revenue difference in October 2019 is negative. I need to understand the accounting issue before treating it as ordinary economic behaviour.
What I have tested so far
My first forecasting experiment asks a focused question: can a model predict next month's reported government revenue better than a simple guess?
I start with two simple comparisons: use last month's revenue, or use revenue from the same month a year earlier. I then test AR(1), ARIMA and seasonal ARIMA. AR(1) uses the previous month's level; ARIMA can model changes and patterns in forecast errors; seasonal ARIMA also includes a yearly pattern in the monthly series. These are fixed starting models, not a final choice.
All five methods predict the same 24 months, from July 2024 to June 2026. For each prediction, the model only gets the earlier fiscal observations. I start with 102 training months. The data comes from the saved, revised source tables, so this is a historical test rather than a reconstruction of exactly what someone knew at each publication date.
I compare how far each prediction was from the reported result. MAE is the average size of the errors. RMSE gives larger mistakes more weight. In this first test, AR(1) has the lowest MAE and ARIMA has the lowest RMSE. The seasonal model performs worse. A more complicated model has not automatically helped.
I have not chosen the final model. I still need to check the remaining patterns in its errors, unusual months and changes in the series. The shaded 95% prediction band shows uncertainty under the model's assumptions. It is not a guarantee. The saved July–September 2026 predictions use data ending in June; they are not a fresh forecast made today.
What the budget scenarios mean
I use the IMF's April 2026 projections through 2031 as a separate yearly starting point. The application lets me change an assumption, such as a 10% fall in oil prices, and calculate how the budget balance changes.
This asks what would happen under these assumptions, rather than saying what will happen. In the simple price scenario, I keep spending, export quantities, other revenue, GDP and the exchange rate fixed. I choose how much revenue is assumed to respond to the price change. That share is visible and adjustable.
The result helps me see how sensitive the budget is to the assumptions. It does not include every government response or economic effect, and it does not assign a probability to the scenario.
What the application does
My Python application brings the sources, calculations and results together. I can explore the yearly history, inspect monthly accounts, compare forecast errors, see prediction uncertainty and export results. I can also compare countries using oil rents, real output per person and manufacturing, then change how much weight each feature gets.
Those country comparisons show measured similarity within my selected set. They do not prove why one country developed differently from another. I want to use them as a starting point for better questions and a closer reading of the evidence.
The website has a lighter browser preview of the saved results. It can redraw the charts, change comparison weights and recalculate the budget scenario. It does not run new Python model fits. The complete Streamlit application and its code stay in my private GitHub repository while I develop the project.
Where I want to take the program
Eventually, I want this program to be useful for any country, with a clear way to import different datasets. For now, I am building and testing it around the Iraqi data I understand and care about. Once that works well, I want to make the import process, country choices and analysis easier to adapt. Different datasets will still need their own checks and suitable methods; that wider version is something I want to develop after the core thesis.
I am keeping that broader aim separate from the minimum thesis. First, I need a reliable application for the questions and data I have already checked.
Why I value my professors' guidance
Professor Uscatu and Professor Davidescu have both agreed to work with me on this project. I respect them greatly, and I look forward to learning from them and building something I can be proud of.
I look forward to Professor Uscatu's guidance on the software: clear modules, sensible storage, reliable algorithms and an application that runs smoothly. I want to understand the code well enough to explain and defend it myself.
I look forward to Professor Davidescu's insight into the data, statistical methods and interpretation. Her guidance will help me ask stronger questions, review the models carefully and understand what the results really mean.
I want their advice to challenge my ideas and help me grow as a programmer and researcher. The title and final scope will be refined with Professor Uscatu as the project develops.
What I want to improve next
Following Professor Davidescu's feedback, I have added a more visible national picture of Iraq, international-source projections and a dashboard bringing the results together. I keep separate April and October 2026 IMF releases. The new annual multivariate growth experiment uses previous GDP growth, a previous price-adjusted spending-growth proxy and previous world oil-price growth. It fits 2006–2025 after transformations and lags, within the wider study starting in 2003. It does not outperform the simple expanding-mean benchmark over eight common test years; I show that result and the uncertainty.
I also investigated the suggested VECM. The reviewed annual level candidate has only 18 complete years, with inconsistent integration screens and sensitive rank results. The quarterly source investigation found substantial date coverage, but unresolved revisions, mapping and valuation differences. I have documented those problems and guarded the VECM interface rather than presenting an unsupported fit. A harmonised quarterly series and a reviewed specification remain the next requirements for that method.
My own explanation idea is now implemented with four questions: what changed, what it means, what to investigate and how certain it is. The dashboard connects reviewed events, declared patterns, account reconciliation, separate spending classifications, the 2021 informal-work survey, price/quantity calculations and one/three/six-month forecast evaluation. This adds useful software behaviour without treating nearby events as causes or accounting discrepancies as theft. My demonstration guide explains the implemented version and its limits.
The foundation is in place: checked data processing, monthly analysis and forecast comparisons, yearly development context, clear scenarios and a working Python application.
Next, I want to review the data warnings and model checks, improve the database and application, and investigate spending categories where comparable records are available. I will keep the thesis focused enough to finish and understand it properly, while leaving room for the program to grow.
I can't wait to see what my teachers cook up for me: ideas to add, questions to explore and ways to make this work better. The sky is the limit for where I hope to take it. I will put all the passion I have into making this a useful tool for anyone interested in understanding a country's economy, just as I am trying to understand my own.
Where the evidence comes from
My annual development data comes from World Bank World Development Indicators. My monthly government accounts come from Central Bank of Iraq statistical bulletins. My yearly outlook comes from the IMF World Economic Outlook, with separate April and October 2026 editions. The labour view uses the Iraq Labour Force Survey 2021, and spending classifications use the IMF's July 2025 Iraq report. Exact file links, pages, units and source checks are kept in the project's data inventory.
For forecasting, I am using Forecasting: Principles and Practice and the official statsmodels documentation. The more detailed method notes remain in my private research folder.
For the 2003 starting point, I checked the US Office of the Historian and the UN Secretary-General’s December 2003 remarks.