Creating a forecast balance sheet

Financial analysis is a demanding exercise: cross-checking several fiscal years, spotting trends, computing ratios and projecting the next year. This use case shows how Hermes Agent, the open-source self-improving agent from Nous Research, can take over the full analysis of your balance sheet. You give it your balance sheets from previous years and it reads them, asks you questions about your activity and your assumptions, then builds a forecast balance sheet for next year delivered as a real-time editable HTML site.

What makes Hermes Agent powerful is that it reads your local files and understands their structure, whether they are PDFs, spreadsheets or accounting exports. It operates from your own machine and your sensitive financial documents never leave your hardware.

Hermes Agent does not replace your accountant. It already helps you do 90% of the work and prepare everything as well and as quickly as possible for your balance sheet.

The goal

Build a single working folder with all your accounting documents and let Hermes Agent:

• Read your balance sheets from previous years before analyzing
• Build the context of your company by asking you precise questions about your activity, your sector and your assumptions
• Compute the indicators and ratios from the data it read
• Generate a forecast balance sheet for next year as an HTML site with cells editable by any user

Hermes, a person you employ

Think of Hermes as a financial analyst you employ for your balance sheet. You explain your need, it takes notes, reads your files and gives you advice. It understands your activity, remembers your assumptions from one session to the next and asks the right questions at the right time. Like a good assistant, it does not just collect the numbers: it checks them, spots inconsistencies between fiscal years and suggests adjustments. This working relationship, close to the one you would have with a human analyst, is what makes the analysis more reliable and faster.

The setup

Create a working folder and put your files in it:

1. balance-sheet-2024.pdf — the balance sheet from the previous year
2. balance-sheet-2023.pdf — the balance sheet from the year before
3. balance-sheet-2022.pdf — a third fiscal year for trends

Hermes generates the CONTEXT.md file itself by asking you questions. This file is the key, it gives Hermes the context it cannot infer from your balance sheets alone and looks like this:

# Company context

Sector: wholesale food
Legal form: LLC
Size: 18 employees
Currency: euro (€)
Assumptions for next year:
- Revenue growth: +6%
- Purchase cost increase: +4%
- New hire: 1 salesperson
- Planned investment: delivery truck

Choosing a model for Hermes with Ollama

Hermes runs against any OpenAI-compatible provider, including a local model served by Ollama. Running an autonomous agent locally on modest hardware means balancing three competing factors: reasoning capability ("thinking"), structured tool execution and inference speed.

Models that support reasoning traces generate extra "thinking" tokens before emitting their action output. On CPU-only systems, a large model produces so many tokens that each tool-calling loop slows down drastically. The fix is a small model that still reasons, like deepseek-r1:1.5b or qwen3:1.7b. It keeps the token count low enough for fast execution on CPU while keeping enough structured output quality for Hermes to call tools reliably.

Pull the model, then let Ollama configure Hermes:

ollama pull deepseek-r1:1.5b
# create the 64k context size model -> deepseek-r1:1.5b-64k
ollama launch hermes
# choose the deepseek-r1:1.5b-64k model by prompting "/models"

The model context size must be above 64k for Hermes to run reliably. Ollama defaults to a much smaller context window, so you need to reconfigure your model with the extended context hack described in Fix raw XML tool calls from Ollama models.

ollama launch hermes asks you which model to use, installs Hermes if needed and sets the local Ollama provider for you. Everything is configured automatically, no server address to type.

You can raise the parameter size later on a more powerful machine; the workflow stays the same.

A local model is also the most private option. Your balance sheets are sensitive financial data. With Ollama serving the model on your own GNU/Linux machine, your documents are read and analysed locally and never leave your hardware. No financial data travels to the internet.

Launching Hermes and ask for the first prompt

Ask Hermes to read the files first, then analyze:

Read balance-sheet-2024.pdf, balance-sheet-2023.pdf and
balance-sheet-2022.pdf from this folder. Build the context of my
company by asking me questions, then generate a forecast balance
sheet for 2025 as an HTML site with cells editable by any user.

Let Hermes ask questions

Hermes will read your files and almost certainly come back with questions. Answer them precisely, because these details are what the forecast balance sheet depends on:

• The sector of activity and the legal form
• The revenue growth assumptions
• The expected cost variations (purchases, payroll, rent)
• Planned investments or loans
• The customer and supplier credit policy

For example, answer with a complete sentence such as:

Wholesale food, LLC of 18 employees. Revenue growth of +6%,
purchase cost increase of +4%. A new salesperson hired in March,
investment in a delivery truck in the second half of the year.

The more specific your answers, the more relevant the forecast balance sheet. Your answers also build and enrich the CONTEXT.md file over the sessions.

Generating the editable HTML site

Once the files are read and the context is built, Hermes generates a standalone HTML site. Each line of the forecast balance sheet is presented in an editable cell: any user can click on a value, edit it directly in the browser and see the total recalculate instantly. No installation or server is needed, the forecast-balance-sheet.html file opens in any browser.

The site contains:

• The assets side and the liabilities side of the balance sheet side by side
• The historical values read from your PDFs
• The forecast values computed by Hermes
• Editable cells for every line item
• The total recalculating live on every change
• An export button to download the edited balance sheet as CSV

Hermes writes all of this into a single HTML file with the necessary JavaScript included, so you can share it by email, drop it on a network share or host it on a small internal server.

Consolidating results over time

Hermes persists every session, so the analysis does not restart from zero each time:

• Resume the conversation with hermes --continue or hermes -c in the same folder
• Ask Hermes to maintain an analysis.md tracker: every assumption is logged with its source (read from a file or entered by you)
• Its memory and skills improve with use, so repeated analyses get faster and more accurate
• For a hands-off approach, schedule a monthly regeneration with hermes cron and review the updated forecast balance sheet when you want

Over a few sessions you get a living forecast balance sheet that grows, corrects itself and remembers every assumption you already tested.

The final verdict

Hermes Agent is comprehensive on every part of the balance sheet analysis: it reads and understands all your accounting files, from previous balance sheets to appendices. The best part is that it behaves like a very talented teammate on the task. It gives you advice, spots the inconsistencies you would have missed and remembers everything you have worked on together. Because it keeps every session and generates an editable HTML site, you get a deliverable that anyone on your team can adjust without technical skill. You are not following a script; you are working with someone who knows your financial situation deeply and follows it through to the end.