AI × FINANCE
Explore AI in Finance
AI can extract information from financial filings, flag payments for investigation, and estimate future quantities. Each task needs a different kind of evaluation.
Examine a Treasury fraud-recovery announcement, learn how to trace a filing metric, and consider a project using historical economic data.
Select a step to see what happens.
Public filing
Locate the reported figure
Keep the entity, financial concept, currency, scale, and reporting period beside the value.
Extraction
Find the passage or structured fact
An assistant can propose a relevant passage. EDGAR financial facts offer another route to data, but matching labels do not guarantee comparable periods.
Calculation check
Recompute before interpreting
Check the formula and original filing. Distinguish a reported fact, a calculated change, and a scenario based on assumptions.
Illustrative filing analysis. The scenarios are not forecasts or investment advice.
THE BASICS
A financial number needs a definition
Finance examines money, financial decisions, and institutions. Economics examines questions about production, incentives, and wider economic conditions. Before analyzing a figure in either field, record its definition, period, unit, and comparison.
Language models can help locate passages in documents. Classification and anomaly detection can prioritize records for investigation. Forecasting models estimate future quantities from observations. An agent may connect retrieval with calculation tools; the source values and calculations still need checks.[1]
A first project can follow one metric through two reporting periods or describe one historical economic series. Keep the source values and calculation visible so another reader can reproduce the result.
APPLICATIONS
How does AI work with financial information?
Follow the information through each task: the input, the model’s contribution, and the person responsible for checking the result.
Reported industry use and public data infrastructure
Finding a metric in a financial filing
A quick summary of a long report can miss the context that makes a number meaningful.
- The material
- A company's public annual or quarterly filing, with its reporting dates and notes.
- AI’s role
- Treasury's 2024 report describes document extraction among reported generative-AI uses. An assistant can propose a relevant passage; the SEC also provides structured XBRL financial facts through EDGAR APIs.
- The human decision
- Confirm the original passage, financial concept, currency, and period. Record whether a figure describes a duration or a point in time.
The limitThe SEC warns that reporting calendars differ. A matching label or calendar frame does not eliminate the need to inspect the underlying dates and context.[1][2]
Operational government example
Prioritizing a suspicious payment for investigation
A large stream of payments makes it difficult to identify which records deserve attention first.
- The material
- Authorized payment records in a controlled institutional setting; an educational example should use synthetic records.
- AI’s role
- Treasury reports using machine learning to expedite identification of Treasury check fraud. A flag can direct attention to a candidate problem.
- The human decision
- Investigate the evidence and decide what action is justified. For a student model, inspect both wrongly flagged legitimate examples and missed synthetic fraud examples.
The limitAn unusual record is not proof of wrongdoing. The Treasury announcement does not provide a complete classification test or a transferable accuracy rate.[3]
Reported financial use; public economic research material
Studying a forecast without hiding uncertainty
An estimate may look convincing even when it uses information unavailable at the time it supposedly predicted the future.
- The material
- A defined historical economic series, its source, frequency, units, and information-release dates.
- AI’s role
- Predictive models can estimate a future value from earlier observations. Treasury describes forecasting among financial-sector uses; FRED and ALFRED make economic observations and historical information vintages accessible.
- The human decision
- Choose a forecast horizon, compare with a simple baseline, and reserve later observations for evaluation. Check what was actually known at each prediction date.
The limitFRED explains that past observations can be revised. Today's version of an old value may introduce hindsight into a historical forecast comparison.[1][4]
CASE STUDY
Read Treasury's FY2024 results by category
The announcement combines prevention and recovery. Machine-learning check-fraud recovery is one component of the total.
On October 17, 2024, Treasury reported over $4 billion in fraud and improper-payment prevention and recovery for fiscal year 2024, October 2023–September 2024. Its breakdown attributed $1 billion in recovery to faster identification of Treasury check fraud using machine learning. Other components covered screening, prioritization, and payment-processing changes.[3]
Prevention and recovery are different outcomes, and the department-wide process included several tools and interventions. The release reports operational results. It does not isolate the model's contribution through a randomized comparison.
In your analysis, label the outcome, amount, time window, and responsible process. Ask for case counts and false alarms before treating a recovery headline as a measure of model accuracy.
Follow the process Open a step to see what happens.
Identify the quantity
Dollars recovered, dollars prevented, or cases flagged?
Keep the period
Which fiscal year or observation window?
Check attribution
Which result belongs to the model, the workflow, or the wider organization?
Keep recovered dollars, prevented payments, flagged cases, and detection accuracy separate in your report. The announcement supplies some of those measures, not all of them.[3]
SKILLS & STUDY
What to learn first
- Read the financial meaning
- Learn revenue, expenses, cash flow, assets, and liabilities. State the metric rather than treating every dollar amount as interchangeable.
- Keep units and denominators
- Write currency, scale, dates, and the denominator for each rate. Percentage change uses a starting value; a percentage-point difference compares two percentages.
- Use reproducible calculations
- Begin with a spreadsheet and an explicit formula. Programming helps automate collection; statistics and time-series knowledge become necessary when you train or evaluate a forecast.
- Separate observation from assumption
- Label reported facts, calculated values, and illustrative scenarios separately. A scenario says what follows from chosen assumptions; it does not establish what will happen.
PROJECT IDEAS
Projects with traceable numbers
Choose a question, inspect the requirements, and keep a brief for later. Each proposal includes a way to check the work.
Save an idea in this browser or copy its brief to discuss with a mentor.
No account neededTrace a filing metric from source to calculation
Can you explain how one comparable metric changed between two reporting periods?
- Intended output
- A source-linked table and short explanation with the formula, unit, dates, and relevant filing note.
Skills, sources and checks
- Starting knowledge
- Basic accounting vocabulary, percentages, and spreadsheets. An API-based extension adds coding and structured-data handling.
- Material to consider
- Public SEC filings and corresponding EDGAR financial facts. Choose one entity before attempting comparisons across companies.
- How to judge the result
- Reconcile extracted values with the filing. Confirm comparable periods and definitions, and have a reader reproduce the calculation.
A reporting analysis focused on one metric and comparable periods. Investment recommendations and valuation are outside the scope.
Explain a historical economic pattern
How does one clearly defined economic series change across a chosen historical interval?
- Intended output
- An annotated chart, a reproducible calculation, and a note distinguishing description from a causal explanation.
Skills, sources and checks
- Starting knowledge
- Graph reading, rates, and data definitions. Forecasting is an optional advanced extension requiring a separate evaluation plan.
- Material to consider
- A public FRED series with its original provider and metadata. Use ALFRED information vintages if the question depends on what was known earlier.
- How to judge the result
- Check frequency, units, transformations, missing observations, and data vintage. Explain how the chosen start and end dates affect the result.
A descriptive historical analysis. A causal explanation or trading rule would require a different study.
Use public filings, economic series, or synthetic payments. These educational analyses do not provide personal investment advice or access financial accounts.
MENTOR SUPPORT
Plan a source-linked finance workflow
A filing-analysis workflow could retrieve a passage, check a calculation, and record the supporting source. Discuss the question, available material, and evaluation with a mentor in the STEMtiq Agentic AI Scholars Program.
Finance interests require a project scope that fits the shared agent-building curriculum and your preparation. Specialist forecasting and financial-model training are separate commitments.
5 weeks · 25 live hours · Grades 9-12
COMMON QUESTIONS
What students ask
Is AI in finance mainly about picking stocks?
No. Financial work also includes extracting information, investigating unusual transactions, and assessing uncertainty. A student can study those tasks using public or synthetic material without making an investment recommendation.[1]
Why can two financial numbers with the same label differ?
Check the entity, currency, reporting dates, scale, and definition. A quarterly amount, annual amount, and point-in-time balance describe different things. The SEC's API guidance specifically notes differences between companies' reporting calendars.[2]
What does a historical data vintage mean?
It is the version of information available at a particular time. An economic observation may be revised later. ALFRED lets researchers retrieve information known in an earlier period, which matters when evaluating a past prediction.[4]
How much math does a first finance project require?
A carefully scoped filing comparison can begin with arithmetic, percentages, and a spreadsheet. Anomaly-model evaluation adds classification measures; forecasting adds statistics and time-series reasoning. Choose a question whose calculations you can explain yourself.
Can AI guarantee a good financial outcome?
No. The joint SEC, NASAA, and FINRA investor alert identifies guaranteed-return claims as warning signs and cautions that generated information can be inaccurate. A plausible explanation is a starting point for verification, not a promise about future results.[5]
Discuss a public-data finance project.
Choose one public filing or economic series and a calculation you want to explain. Discuss the data, required skills, and evaluation with a mentor.
Sources & further reading
- [1]
- [2]
- [3]
- [4]Federal Reserve Bank of St. Louis: FRED and ALFRED: Real-Time Periods (Undated technical documentation)Return to guide ↑
- [5]
Sources inform this guide; their organizations do not endorse STEMtiq. Project ideas and diagrams are original editorial explanations.