Valuation models with Deepseek v3 and Excel Sheets
re you feeling innovative today? With Excel? Would you want an AI to do some Excel magic for you? And that with full control over the data? Then read the following article.
The following information, which I am about to share, isn’t intended to be taken as financial advice of any kind, and it should not be used to support or guide any financial decisions you might consider. These insights represent my independent opinion, formed from my specific interests in the fields of finance and technology.
Fintech changes the world of quantitative finance. One spread sheet at a time.
Models are the bread and butter of quantitative finance that professionals rely on. These include various valuation models, detailed business plans, and diverse scenarios, along with critical financial performance indicators. Additionally, one cannot overlook the importance of meso- and macro-economic fundamentals in shaping our understanding of financial markets.
Investment bankers and traders spend a lot of time with their models. And much of this work is done with Excel. Occasionally with R or Python if you are a Quant (a Quantitative Analyst, usually in finance specifically). But Excel is a standard because Microsoft has developed it since 1987.
I was negatively surprised when I uploaded a standard Excel spreadsheet (MS-XLS without Macros) to OpenAI (GPT-4) and Anthropic (Sonnet) services. The models cannot fully read Excel tables. They also don’t understand the different types of Cash Flows. If you don’t distinguish between Free Cash Flow and Discounted Cash Flow, your price predictions will not be good, to say the least. Still, you hear about financial influencers we claim to be profitable with this kind of technology. I have also observed that it’s very difficult to spot their mathematical errors. Especially after a long day of work.
What is DeepSeek?
DeepSeek is the competition from Asia, a Chinese startup in the Chinese Artificial Intelligence industry (Hangzhou Deep Search Artificial Intelligence Basic Technology Research Co., Ltd.). For now, this company is focused on research, and the work is Open-Source. This means that you can self-host this AI, and run it on your own (high-end) hardware. As a tech-savvy Fintech engineer, you could support your peers with hosting it on popular AI hubs, in a private colocation data-center or in popular on-demand public-cloud services like AWS (Bedrock). The possibilities of Open-Source AI for Fintech is especially great, given that you don’t need to share your proprietary in-house concepts and data with arbitrary third-parties. Meaning: even though this is a sophisticated Chinese model, the data is not exported to China, ensuring full compliance with privacy standards. You can take complete control over the data, manage its usage effectively, and still archive the same impressive productivity benefits that you expect and deserve.
DeepSeek v3 is the latest and greatest version now (December 2024).
Excel and models
DeepSeek v3 is capable of reading Excel spreadsheets within its chat window. This works excellent. I uploaded a private valuation model (German) to their hosted service to test the capabilities.
DeepSeek v3 gives concrete advice on an Excel document with various cells, sheets, multiple tables and quantitative financial approaches.
Can’t Microsoft Copilot do this?
When Microsoft acquired OpenAI, I mean when they formed their AI partnership of course, many people in the industry expected that flagship products like MS 365 would benefit from this. But the fact of the matter is that Microsoft’s Copilot is a joke. It can’t even read basic tables, or perform the most basic actions. Its reasoning capabilities are substandard, and the costs are very high.
The answer is: No. Microsoft’s Copilot AI isn’t even 10% as capable.
What are the concerns regarding DeepSeek?
DeepSeek Cloud is specially trained with its output, intricately designed based on the user data it receives. For those operating within the finance domain, I would advise strongly against utilizing the cloud-hosted service due to the inherent risks and potential vulnerabilities associated with such an approach.
DeepSeek was trained (to some degree) with OpenAI GPT4o output. There may be some legal issues here.
DeepSeek uses more than 600B parameters, which means that hosting it is expensive. High-end models such as Llama (from Meta) have around 400B. B stands for Billions here, like the TV series ;) In other words: this is expensive.
It’s possible to address all of these concerns with self-hosting. You can also trim down the costs intelligently, I believe. That would be an interesting project.
Summary
DeepSeek (v3) offers Open-Source coding assistants to the software developers. We all know that software is changing the finance industry. We also know that AI is going to make us more productive within Excel, valuation models, and market research now. Even with software from the 1980ies. What sticks out here is, that the performance of the Open-Source model is higher than of the proprietary competition from the USA. And that this affects the bread and butter of quantitative finance.
Every boutique investment bank or small trading company can self-host DeepSeek and improve their strategies. You don’t need Bloomberg AI or a multi-million dollar setup. That means even retail traders will at some point use this such AI systems.
Perspectively, we can see that Excel has gotten Python support recently. Which means it’s only a matter of time until AI with intuition and reasoning capabilities will allow us to integrate more abstract business model data. This can help with scenario-based analysis, or even with backtesting securities trading strategies.
