Hi - good stuff, I've been doing a more or less manual version of this with Excel in Claude and NotebookLM. One thing I found (and have found to be the case with your prompt that tries to download PDF 10Qs, transcripts etc) is that it often doesn't work at all - and for example it didn't work for me now having tried it in Cowork for Coinbase. I had a prompt that worked for some stocks but not for others - in terms of downloading PDF verions of financials. I'm wondering if you are better able to extract data with Claude Code's more tailored approach v's Cowork? Does that really work for you all the time? Regardless, my 2c is that it's just more reliable and just as fast to do the data downloads from the IR site and download everything there - the EDGAR site is not good esp when you need PDFs (as NotebookLM needs). That said I would love to have every new filing and transcript automatically downloaded into a directory and them automatically uploaded to NBLM etc - but guessing I need to wait a while for this to become easy.
I have a script from Claude Code that extracts all filings from the EDGAR API, but not in pdf format. LLMS don't like pdfs. Need to check but I believe it's XML format. When I ran this on 100 saas companies, it pulled everything. If it does not work the first time in cowork, the best way is to ask Claude to adapt the script. Once the edge cases are resolved, it should pull everything. I agree Claude Code remains the most versatile one, but it does require a bit more knowledge.
Thx. I use Claude and Gemini, and Gemini being more multimodel takes PDFs easily. NotebookLM actually requires PDFs rather than XML. Come to think of it Claude in Excel accepts PDFs - when model building you just upload a single 10k and the last 3 10Qs (found that it's best not to overload it with stuff including prompt size fyi).
Thanks for the feedback. I do tend to separate the financial data from the annual reports. Financial data is best in CSV format, so I use yfinance or a paid data provider. The annual reports include the text and the notes. I don't, for example, extract the income statement from the annual report.
I didn't run into that problem. After building the skill, I ran it as a test in Cowork. It pulled 5 years of data for 3 companies and gave me an Excel with an assessment for each. Also, the skill does not ingest all the data of the annual report. It looks at specific sections of it, so that limit context creep.
What I like about Claude is that you can build almost anything with it. Give Claude the idea, then say: "Ask me questions." What do you need from me to get this to work?
Same thing with valuation models and company analysis. If you can do the work up front and provide a solid framework, you can start iterating to find quality. Most people seem to give up after an hour, but even with AI, there is no free lunch. We need to do the work. The nature of the work has changed. Thanks for your feedback!
Honestly, quants have been using AI for a long time to help them make buy/sell decisions. I believe a quant strategy in Portfolio123 with AI assistance will outperform what Claude can do on its own. P123 has high-quality time-series data. What data are you feeding Claude?
Hi - good stuff, I've been doing a more or less manual version of this with Excel in Claude and NotebookLM. One thing I found (and have found to be the case with your prompt that tries to download PDF 10Qs, transcripts etc) is that it often doesn't work at all - and for example it didn't work for me now having tried it in Cowork for Coinbase. I had a prompt that worked for some stocks but not for others - in terms of downloading PDF verions of financials. I'm wondering if you are better able to extract data with Claude Code's more tailored approach v's Cowork? Does that really work for you all the time? Regardless, my 2c is that it's just more reliable and just as fast to do the data downloads from the IR site and download everything there - the EDGAR site is not good esp when you need PDFs (as NotebookLM needs). That said I would love to have every new filing and transcript automatically downloaded into a directory and them automatically uploaded to NBLM etc - but guessing I need to wait a while for this to become easy.
I have a script from Claude Code that extracts all filings from the EDGAR API, but not in pdf format. LLMS don't like pdfs. Need to check but I believe it's XML format. When I ran this on 100 saas companies, it pulled everything. If it does not work the first time in cowork, the best way is to ask Claude to adapt the script. Once the edge cases are resolved, it should pull everything. I agree Claude Code remains the most versatile one, but it does require a bit more knowledge.
Thx. I use Claude and Gemini, and Gemini being more multimodel takes PDFs easily. NotebookLM actually requires PDFs rather than XML. Come to think of it Claude in Excel accepts PDFs - when model building you just upload a single 10k and the last 3 10Qs (found that it's best not to overload it with stuff including prompt size fyi).
Thanks for the feedback. I do tend to separate the financial data from the annual reports. Financial data is best in CSV format, so I use yfinance or a paid data provider. The annual reports include the text and the notes. I don't, for example, extract the income statement from the annual report.
Hey, i like the logic. One question. How do you solve the context window limit? With that many documents, you will run out of context on Sonnet.
I didn't run into that problem. After building the skill, I ran it as a test in Cowork. It pulled 5 years of data for 3 companies and gave me an Excel with an assessment for each. Also, the skill does not ingest all the data of the annual report. It looks at specific sections of it, so that limit context creep.
Great work on this — people don’t understand how efficient Cowork is.
I have been more successful with feeding cowork my templates such as a DCF.
Have to build it upfront, but most serious investors have this built anyway
What I like about Claude is that you can build almost anything with it. Give Claude the idea, then say: "Ask me questions." What do you need from me to get this to work?
Same thing with valuation models and company analysis. If you can do the work up front and provide a solid framework, you can start iterating to find quality. Most people seem to give up after an hour, but even with AI, there is no free lunch. We need to do the work. The nature of the work has changed. Thanks for your feedback!
Appreciate the response Kevin, and certainly! Will let you know if I have any tricks up my sleeve or find anything interesting going forward
Nice. Doesn't replace the hard work, though. But I do believe we can do more now with the same time we have.
Are you the same as Claude Portfolio, over on X? https://x.com/claudeportfolio
Honestly, quants have been using AI for a long time to help them make buy/sell decisions. I believe a quant strategy in Portfolio123 with AI assistance will outperform what Claude can do on its own. P123 has high-quality time-series data. What data are you feeding Claude?