Goodbye Excel: Building a Financial Portfolio App with AI Coding Tool Codex

From Excel to AI Coding: A New Option for Finance Professionals

For decades, Excel has been the go-to tool for finance professionals building applications. It's simple, fast, and intuitive — when a business unit needs a new app, they often can't wait three to six months for the IT team's development cycle, so they build a first version in Excel. The spreadsheet may not be robust or enterprise-grade, but it works and keeps the business moving forward.

Since its release in 1985, Excel has become the de facto standard tool for the global financial industry. It's estimated that over 750 million people use Excel worldwide, with nearly 100% penetration in finance. From investment banking DCF valuation models and hedge fund risk analysis to corporate financial budgeting, Excel is virtually everywhere. JPMorgan Chase once disclosed that it runs hundreds of thousands of mission-critical Excel files internally. However, this dependence also carries enormous risk — during JPMorgan's 2012 "London Whale" incident, an Excel formula error contributed to $6.2 billion in trading losses. Excel's "shadow IT" problem (business units building systems that bypass IT departments) has long been a pain point for compliance and risk management at financial institutions.

Yet this decades-old workflow is being fundamentally transformed by AI coding tools. A seasoned financial technology professional shared his experience on Bilibili: Using the AI coding tool Codex, he built a working stock portfolio management tool from nothing more than a natural language description.

Excel's Limitations and Where AI Coding Tools Fit In

What Excel Got Right

Excel's core value lies in helping finance professionals "think in models." It enables non-technical users to quickly translate business logic into computable models — and that remains important to this day.

Where AI Coding Tools Are Best Suited

The video's author clearly defined the boundaries for AI coding tools:

• Good fit: Personal tools, prototype validation, internal experiments, small productivity apps
• Not a good fit: Mission-critical enterprise systems — these still require professional development, testing, security audits, governance, and production support

This positioning is highly pragmatic. AI coding tools aren't meant to replace professional development teams — they fill the massive gap between "Excel prototypes" and "enterprise software." In the past, turning a finance professional's idea into usable software required lengthy requirements gathering and development cycles. Now, if you can clearly describe your financial workflow, AI can turn it into a working application.

Hands-On Demo: Building a Portfolio Management Tool from Scratch

One Prompt, Codex Does the Rest

The author used OpenAI's Codex for the demonstration, describing a simple but complete requirement to the AI:

Build a US equity portfolio management application that can download security master data, support trade entry, and perform end-of-day valuation.

Codex is OpenAI's AI coding agent, built on its latest large language models. Unlike traditional code completion tools, Codex can autonomously execute complete software development tasks in a cloud sandbox environment — including reading codebases, writing code, running tests, and fixing bugs — all without human intervention. It works similarly to a junior developer: after receiving natural language instructions, it formulates a plan, implements step by step, and automatically debugs when issues arise. This "agentic coding" paradigm represents the latest evolution in AI-assisted development, progressing from "humans write code, AI completes it" to "humans describe requirements, AI handles the entire development."

With just that single natural language description, Codex began working autonomously. Throughout the process, it displayed its reasoning and current steps, even automatically fixing errors and retrying when they occurred.

The Result: A Fully Functional Web Application

Within minutes, Codex generated a web application accessible via URL, containing the following functional modules:

1. Security Master

• Automatically downloaded data for 12,599 securities from NASDAQ
• Including 7,390 stocks and 5,209 ETFs
• Complete data with clear categorization

Security master data is one of the foundational pieces of infrastructure for financial institution operations. It contains each security's identifiers (such as CUSIP, ISIN, Ticker), classification information, exchange affiliations, corporate action data, and more. At large asset management firms, maintaining an accurate security master typically requires dedicated teams and expensive data vendors (such as Bloomberg and Refinitiv). Data quality issues — like ticker changes and mapping errors caused by corporate mergers — are among the most common sources of failures in investment operations. While Codex's ability to automatically download and categorize nearly 13,000 securities from NASDAQ would still need more rigorous data governance in a production environment, it already demonstrates AI tools' capability to rapidly build data pipelines as a proof of concept.

2. Trades

• Supports entering ticker symbols, buy/sell direction, quantity, and price
• Each trade is automatically appended to the Trade Ledger

3. Portfolio & EOD Valuation

• Automatically calculates position cost basis ($20,000 in the demo)
• One-click end-of-day valuation that downloads the latest closing prices
• Automatically calculates current portfolio market value and P&L

End-of-day (EOD) valuation is a core process that must be executed at the close of every trading day in the asset management industry. The basic logic is: obtain the latest closing prices for all securities held in the portfolio, multiply by position quantities to get market value, then compare against cost basis to calculate P&L. In actual operations, this seemingly simple process involves complex pricing hierarchies (exchange prices, broker quotes, model pricing), multi-currency conversions, accrued interest calculations (for bonds), and final NAV (Net Asset Value) confirmation. At large fund companies, the EOD valuation process is typically handled by dedicated investment operations teams using specialized systems like Charles River or SimCorp. While the AI-generated version simplifies these complexities, the core logical framework is correct.

This isn't a simple calculator — it's a complete financial application with data acquisition, trade management, and valuation capabilities — and its entire "development process" was just writing one paragraph.

Deeper Implications: A Paradigm Shift in Financial Workflows

From "Model Thinking" to "Software Thinking"

The author offered an incisive summary:

Excel helps us think in models; AI coding tools help us turn models into software.

This statement reveals an important paradigm shift in the financial industry. Previously, finance professionals' capabilities stopped at the Excel model; now, the same business knowledge can be directly transformed into working software applications. This means:

• Dramatically faster prototype validation: Going from idea to demonstrable tool may take minutes instead of weeks
• Significantly reduced communication costs: Rather than writing lengthy requirements documents for development teams, you can first generate a working prototype with AI
• Redefined value of finance professionals: The core competitive advantage is no longer "can you code" but "can you clearly describe business logic"

AI Coding Tool Choices: Codex Isn't the Only Option

You may not have noticed, but the author specifically emphasized that "you can use any coding tool and model to do the same thing." Codex was just one choice for the demo — tools like Cursor, Windsurf, Claude, and others on the market all have similar capabilities. The key isn't which tool you choose, but mastering the new skill of "driving software development with natural language."

The AI coding tools market is currently in a phase of rapid evolution. Cursor is built on VS Code, achieving code editing, multi-file refactoring, and conversational development through deep AI integration. Windsurf (formerly Codeium) emphasizes a "flow-state" development experience where AI proactively senses developer intent. Claude (Anthropic) has an Artifacts feature that can directly generate interactive application prototypes. GitHub Copilot leverages the Microsoft ecosystem to maintain the largest user base. Additionally, tools like Replit, Bolt.new, and v0.dev lean more toward no-code/low-code, enabling users with zero programming experience to generate applications. The common trend across these tools is evolution from "assisting with code writing" to "agentic development," where the user's role shifts from "coder" to "requirements describer and quality reviewer."

Rationally Embracing AI Coding — Not Blindly Replacing Excel

What's most admirable about this demo is its restraint. The author didn't hype "AI can replace everything" but clearly defined the boundaries: personal tools and prototypes are fair game for experimentation, while enterprise systems still need professional teams.

For finance professionals, now is an excellent time to learn AI coding tools. Not to become programmers, but to more efficiently transform years of accumulated business knowledge into usable tools. While your colleagues are still cobbling together complex VBA macros in Excel, you can generate a complete web application with a single paragraph — that's the productivity gap AI coding tools create.

VBA (Visual Basic for Applications) is Excel's built-in programming language and has been the primary means for finance professionals to automate work since its introduction in 1993. However, VBA has numerous limitations: outdated syntax, difficult debugging, inability to easily create modern web interfaces, difficulty integrating external APIs and data sources, and extremely high code maintenance costs. More critically, VBA's learning curve remains steep for non-technical users — understanding loops, conditional logic, object models, and other programming concepts takes considerable time. The revolutionary aspect of AI coding tools is that they completely eliminate this barrier: users don't need to understand any programming concepts; they simply describe business requirements in natural language, and AI generates complete applications written in modern languages like Python and JavaScript, complete with web interfaces. This is a qualitative leap from "learning a programming language" to "programming in your native language."

Key Takeaways

• AI coding tools fill the gap between Excel prototypes and enterprise software, well-suited for personal tools, prototype validation, and internal experiments
• In the demo, Codex automatically built a complete portfolio management application — including security master data download, trade entry, and EOD valuation — from a single natural language description
• Core paradigm shift: Excel helps you think in models; AI coding tools help you turn models into working software
• Finance professionals' core competitive advantage is shifting from technical skills to the ability to clearly articulate business logic
• The author rationally draws boundaries: mission-critical enterprise systems still require professional development teams — AI coding tools augment rather than replace