Expense Manager

expense-manager is an LLM-powered utility that helps you consolidate and categorize all your financial transactions and get a clear picture of where your money comes from, and where it goes 💸

Contents

• What is this?
• What can it do for me?
• How does it work?
• How can I use it?
• What else should I know?
• License

What is this?

expense-manager is an LLM-powered utility that helps you consolidate and categorize all your financial transactions and get a clear picture of where your money comes from, and where it goes. To use it, you just need those transactions in .csv files, which you can typically download directly from your different financial accounts (see an example here)

Next, you pass those files to the expense-manager, which automatically consolidates them into a single view, and assigns a category to each of their transactions choosing from a common, standardized category list (which you could customize as needed).

What can it do for me?

You want to know if you should keep reading or leave now; I'll help you decide in three paragraphs:

ℹ️ Context: Knowing exactly how much money you are making and spending (in each category) increases your financial awareness and helps you gradually refine/improve your decision-making abilities. This contributes to your meta-goal in life: slowly but surely steer reality towards outcomes ranking higher in your personal preferences.

🚩 Problem: Your financial life has become quite complex, and you use multiple banking accounts and credit cards over any given period. You can easily download the activity/transactions for any of them, but the .csv files you get are in slightly different formats, and they either don't categorize your expenses, or they do but use different "buckets" and hierarchies. Consolidating all this information manually is time-consuming and error-prone, and you end up not doing it for months at a time or at all.

🔥 Solution: With expense-manager, you can simply download the transactions for the period you want to analyze (last month, year-to-date, last year, etc.), drop the files in the input folder, run a script, and voilà! you get a beautiful consolidated list with all the expenses categorized consistently. This output will look like this and will be in .csv format, allowing you to further analyze it using a tool like Excel (see the What else should I know? to see how).

How does it work?

This sounds interesting and you want to know more right? You've come to the right section:

Prerequisites

• Python 3
• OpenAI API key added to your environment variables (instructions)
• These dependencies will be installed by the setup script:
  • openai, langchain, tenacity, pydantic: used to make and validate LLM API calls
  • rapidfuzz: used to find similar descriptions that have been categorized in the past
  • python-dotenv: used to load environment variables; you know this one
  • dateparser: date standardization

Process outline:

1. Read all (.csv) files from the input folder and extract the key information required from each transaction (date, type, description, amount).

2. Consolidate all transactions into a single list; this typically involves some level of data wrangling.

3. Transaction descriptions are sent in batches to an OpenAI LLM (gpt-3.5-turbo), which returns the appropriate category for each transaction; the category list can be customized -- see the What else should I know? section for details.

4. The new description-category pairs obtained from the LLM are stored into a reference file. In all future runs, the utility first looks up each description in this reference file and if it finds one that is similar enough, it picks up its associated category and it moves on to the next transaction; this materially reduces the number of API calls required over time, since many transactions are repetitive (we are creatures of habit).

5. The final list of categorized transactions is saved into a .csv file, which you can easily analyze using your tool of choice.

NOTE: expense-manager fully respects your privacy and doesn't collect any data at all. However, it does send the transaction descriptions (and only the descriptions) to an OpenAI LLM to obtain the associated category. OpenAI claims not to sell or use this data for any purposes, (starting on March 1st, 2023, not even to train their models) but just be aware that's how the program works.

Cost estimates

The OpenAI API cost per run will depend on two factors:

1. Total number of calls required
2. Cost per call, which will depend on the number of tokens per call and the price per token (established by OpenAI)

As a point of reference, I've been using this for my own purposes with data from around 10+ different US institutions, and the average call has required around 1300 tokens. This includes both the prompt and the completion required to process 10 transactions since we send 10 descriptions to the LLM in each call.

This means that, in theory, the cost of processing a month of transactions will be around 6 to 13 cents (assuming 250 - 500 transactions per month; see table below for details). In practice, the cost should be lower, since expense-manager won't invoke the LLM if it has seen a similar transaction before (go back to the Process outline section for details)

NOTE: These estimates are based on the pricing of the gpt-3.5-turbo model as of July 1st, 2023 ($0.002 / 1K tokens)

How can I use it?

At this point, you are sold and want to use expense-manager. Here's how to do it:

Installation

Step 1: Clone the repository and cd into the expense-manager folder:

git clone https://github.com/pablovazquezg/expense_manager.git
cd expense-manager

Step 2: Execute the setup.sh script:

./setup.sh

Usage

Step 1: Drop your .csv file(s) in the /data/tx_data/input folder

Step 2: Run expense-manager:

python expense-manager.py

By default, expense-manager will append its output to a master file, effectively creating a historical view of all your transactions. To create a new file instead, use the -n flag (n as in 'new'):

python expense-manager.py -n

➡️ The output file is saved to the /data/tx_data/output folder

By default, input files will be moved from /data/tx_data/input to /data/tx_data/archive after each execution; if you prefer to delete them instead, you can do that using the -d flag (as in 'delete'). Please note this will also delete files processed in previous executions.

python expense-manager.py -d

Last but not least, you can combine the -n (create new file) and the -d (delete / don't archive) flags

python expense-manager.py -nd

What else should I know?

• You can use the output you receive from expense-manager to create a nice income/expense tracker that puts you back in charge of your finances (example here). If you decide to do this, I suggest you split the view between Credits and Debits; you can also color code your expenses to obtain something like this. (some amounts hidden; don't expect totals to match).
• The description-category pairs obtained from the LLM are stored in /data/ref_data/ref_master_data.csv; you can update this list to determine the category to be associated with each description in the future
• If you want to update the income/expense categories (or their associated keywords), you can do that in the <categories> section of the /src/templates.py file
• expense-manager automatically detects and supports American (1,234.56) and European amount formats (1.234,56), as well as many different date formats
• Errors (if any) will be logged in the /logs folder

License