🎯 Key Takeaways for quick navigation: 00:00 📊 GPT-4 can perform financial statement analysis. 01:40 📈 GPT-4 can assess risks and calculate RAC scores based on balance sheet data. 03:55 📉 Interest rate risk is identified as the highest threat to a bank's financial solvency. 04:53 💼 GPT-4's analysis suggests a bank's relative stability but not absolute certainty. 08:48 ⚖️ GPT-4 provides balanced insights on a company's financial position and potential risks. Made with HARPA AI
How well did it use the information your provided (balance sheet) to tweak its response? The responses are general truths rather than analysis that takes advantage of the information provided.
I have done a new test. This time I have used the free version of chatGPT with the AIPRM chrome extension. The result was very similar to the one generated by the GPT-4 chat in your video.
Given what we know about how GPT works (meaning next word prediction), I wouldn't trust this without it explicitly using the Wolfram Alpha beta plugin. It's hard to say whether mathematical knowledge could be captured by the word/text associations contained in GPT's model of the world. For example, is knowing that the characters "1 + 1 =" tend to be followed by "2" the same as actually having a function that manipulates bits to calculate this? I would guess that with more complex calculations, word/character associations are probably less likely to accurately map to the actual output you'd get from a calculator like Wolfram. If it leverages plugins (like calculators and other data stores) and is double- checked by someone with finance expertise, then I think this use case definitely has some potential.
I agree with everything you said here. Instead of just a plugin, If we can add the computational layer like Wolfram as feedback in its learning phase then we might move several leaps ahead. Someone from CBS-Morning interviewed Geoffrey Hinton(father of deep learning) yesterday, and I think whatever he said about ChatGpt like large language models having been ingrained with some degree of understanding and meaning makes complete sense. Consider the following example "Trophy can't fit inside the bag because it's too small" You know this"it" refers to the Bag. Now if I had said "Trophy can't fit inside the bag because it's too large" Now this"it" refers to the trophy instead. Now, When these models are used to translate this sentence from English to French or Hindi, that has pronouns associated with these objects, it still gets it correct. These underlying patterns that govern such choices is what "understanding" is in its essence. From explaining unheard jokes, to designing experiments like "imagine two metal bearings shot from opposite ends of the pipe before the ends are sealed" it "understands" the context and underlying meanings. Some amount of complexity that is required to form such understanding seems to have emerged in these large models.
Actually it is thought that maybe the human brain is also nothing more than a sophisticated prediction machine combined with reinforcement learning. You might argue that next word prediction is just the statistics of the words but when you think that through, what is the word statistics? It is what people tend use? What brings people too use thoese words in the context? That is the model of the world that people have, How to replicate what people would use as a next word? Build a model of the people, infere the model of the world that people use and than come up with the next best word. By doing the above GPT is basically building a model of the world through the written word of human. That model of the world building could be considered intelligence given some high degree of complexity and validity that the model ist generating.
Data you share may be not correct and not contain amount of information which required to take correct position. But however this will help us take decision. I am usng also this to get more clarity form last few months.
Hello. I found your video very helpful for my pdf processing needs using OpenAI. I tried similar approach with hugging Face's language models, but the results are often inaccurate. I'm unsure if it's a model or embeddings issue. I'm thinking of leaving a comment to get advice on improving results with hugging Face's method. Any help would be appreciated. Thankyou!!😊
Thank you for the information. I've tried the same question in chatGPT of Bing. Previously, I made some adoption on the way the chat of BING understand the I was copying the Balance Sheet as CSV file. The answer I obtained it was appropiated but the Bing chatGPT couldn't provide me the RAC score because rhe chat of Bing said it is not familiar with it.
I did similar and also with credit suisse then asked about who’s next on needing a bailout… it suggested Deutsche(but I hadn’t provide any financial information on db).
@@1littlecoder yeah, i shoved in just financial data from svb, and credit suisse, and said these two banks have been bailed out... who's next. i should have thought to shove in db's after it suggested they were next and see if that was still it's thought... after db it was saying some bank in italy.. then paribas... we'll see how things shape up lol.
I don’t think we’re there yet. At this point Chat GP is nothing more than a better calculator or assistant to put information together. I’m emphasizing “at this point” because at some point in the future things might be different but for the foreseeable future we financial accountants will still be indispensable.
Be very careful doing balance sheet analysis with generative AI. It can often make mistakes with simple mathematics. I am talking arithmetic, not even calculus.
Have you tried to add the profit&loss statement to the balance sheet input. Would that add more insight for gpt4. And in the svb case what would happen if you add specific duration mismatch between asset and liabilities? Very busy interesting.
I tried with GTP 3.5. Interesting result. Yes, GTP understooth the Income Statement provided and the most revelant lines (revenue, direct cost, overhaeds and EBITDA) and GTP suggested the direct costs were too high compared with revenue. :) Imteresting.. It was true. But, I tried the same with other companies, and I did not get the same detailed analysis. Maybe with GTP4.... Interesting, promissing, two simple/basc compared with a real human (experienced financial analyst). But helpful when someone do not have any knlewlege about financial analysis, I would say. Maybe helpful to analyze hundersds and hundreds of companies very quikly.
hi, I am ACCA from Pakistan. I can do analysis of financial statements. I would say that GPT4 is good, it is even better than many people I know. Though it is not domain expert (I would not take it's guidance against an ACCA member who has worked in industry for over 5 years) I would still ask it to look out for dangers while making my own assessment. Also, your assessment of poor financial strength from a very old balance sheet is incorrect, GPT4 managed to give better assessment. The reason bank is in finical distress can't be assessed from this very OLD BS i.e. there might have been many other things during that time that lead to this situation.
it's done a good job despite the assessment that the bank is in good health. because it was indeed in good health, and then there came the anti-crypto politicians who ruin the banks crypto environment, and then there was a bank run, and finally, the bank went under.
Financial analysis involves substantially more than a simple balance sheet. The sort of problems which led to SVB failure is not in explicit in the balance sheet. You need to understand the matching of assets and liabilities with respect to duration and timing of cash flows. May it is in the notes to the financial statements, may only management and the Fed has that information…. In short you are feeding insufucient information to GPT4 and expecting it to find things beyond what is possible with what you gave it…
While an interesting exercise, for the most part I don't see GPT-4 adding any value here. For example, it's recommendations are pretty standard actions any competent finance student would say regarding methods of mitigating risk. Impact factors are also straightforward calculations based on the balance sheet. What I find somewhat interesting is the likelihood factor and how GPT-4 is coming up with that number from its training data.
that was a terrible balance sheet analysis - it said nothing incisive. bland, obviously written by a bot with no skin in the game. didn't even compare it against other Banks. well done video info - good test.
Agreed, this kind of analysis is only good for those who are not in the financial analyst world. I mean yeah it looks very good and comprehensive at first. But still far away from what a proper financial analyst could produce.
Sorry its not detailed. Its superficery. What you can check is their Debt, leverage , gear ratios, Dupont analysis, RoA, vs RoCe, etc, these all things determine whether balance sheet going in right direction or not and comlany makeing enough cash to pay its payables. Thanks
🎯 Key Takeaways for quick navigation:
00:00 📊 GPT-4 can perform financial statement analysis.
01:40 📈 GPT-4 can assess risks and calculate RAC scores based on balance sheet data.
03:55 📉 Interest rate risk is identified as the highest threat to a bank's financial solvency.
04:53 💼 GPT-4's analysis suggests a bank's relative stability but not absolute certainty.
08:48 ⚖️ GPT-4 provides balanced insights on a company's financial position and potential risks.
Made with HARPA AI
Wow, that’s incredible! I have a lot of training as a financial analyst and will be using this.
Excellent job!!!
Thank you! Cheers!
How well did it use the information your provided (balance sheet) to tweak its response? The responses are general truths rather than analysis that takes advantage of the information provided.
I have done a new test. This time I have used the free version of chatGPT with the AIPRM chrome extension. The result was very similar to the one generated by the GPT-4 chat in your video.
Thanks for sharing your research! Have a good day!
Please make a video about EdgeGPT
Excellent presentation. Thank you.
Given what we know about how GPT works (meaning next word prediction), I wouldn't trust this without it explicitly using the Wolfram Alpha beta plugin.
It's hard to say whether mathematical knowledge could be captured by the word/text associations contained in GPT's model of the world. For example, is knowing that the characters "1 + 1 =" tend to be followed by "2" the same as actually having a function that manipulates bits to calculate this?
I would guess that with more complex calculations, word/character associations are probably less likely to accurately map to the actual output you'd get from a calculator like Wolfram.
If it leverages plugins (like calculators and other data stores) and is double- checked by someone with finance expertise, then I think this use case definitely has some potential.
I agree with everything you said here. Instead of just a plugin,
If we can add the computational layer like Wolfram as feedback in its learning phase then we might move several leaps ahead.
Someone from CBS-Morning interviewed Geoffrey Hinton(father of deep learning) yesterday, and I think whatever he said about ChatGpt like large language models having been ingrained with some degree of understanding and meaning makes complete sense.
Consider the following example
"Trophy can't fit inside the bag because it's too small"
You know this"it" refers to the Bag.
Now if I had said
"Trophy can't fit inside the bag because it's too large"
Now this"it" refers to the trophy instead.
Now,
When these models are used to translate this sentence from English to French or Hindi, that has pronouns associated with these objects, it still gets it correct.
These underlying patterns that govern such choices is what "understanding" is in its essence.
From explaining unheard jokes, to designing experiments like "imagine two metal bearings shot from opposite ends of the pipe before the ends are sealed" it "understands" the context and underlying meanings.
Some amount of complexity that is required to form such understanding seems to have emerged in these large models.
Actually it is thought that maybe the human brain is also nothing more than a sophisticated prediction machine combined with reinforcement learning. You might argue that next word prediction is just the statistics of the words but when you think that through, what is the word statistics? It is what people tend use? What brings people too use thoese words in the context? That is the model of the world that people have, How to replicate what people would use as a next word? Build a model of the people, infere the model of the world that people use and than come up with the next best word. By doing the above GPT is basically building a model of the world through the written word of human. That model of the world building could be considered intelligence given some high degree of complexity and validity that the model ist generating.
Data you share may be not correct and not contain amount of information which required to take correct position. But however this will help us take decision. I am usng also this to get more clarity form last few months.
Thanks for sharing
Hello. I found your video very helpful for my pdf processing needs using OpenAI. I tried similar approach with hugging Face's language models, but the results are often inaccurate. I'm unsure if it's a model or embeddings issue. I'm thinking of leaving a comment to get advice on improving results with hugging Face's method. Any help would be appreciated. Thankyou!!😊
It's the model issue. The model isn't that good enough. You could try with the recent llama 2 ones. Might give you better result.
Thank you for the information. I've tried the same question in chatGPT of Bing. Previously, I made some adoption on the way the chat of BING understand the I was copying the Balance Sheet as CSV file. The answer I obtained it was appropiated but the Bing chatGPT couldn't provide me the RAC score because rhe chat of Bing said it is not familiar with it.
I did similar and also with credit suisse then asked about who’s next on needing a bailout… it suggested Deutsche(but I hadn’t provide any financial information on db).
Ooh. But ChatGPT data is stuck at 2021
@@1littlecoder yeah, i shoved in just financial data from svb, and credit suisse, and said these two banks have been bailed out... who's next. i should have thought to shove in db's after it suggested they were next and see if that was still it's thought... after db it was saying some bank in italy.. then paribas... we'll see how things shape up lol.
I don’t think we’re there yet. At this point Chat GP is nothing more than a better calculator or assistant to put information together. I’m emphasizing “at this point” because at some point in the future things might be different but for the foreseeable future we financial accountants will still be indispensable.
cool thanks
great video
Yes
Be very careful doing balance sheet analysis with generative AI. It can often make mistakes with simple mathematics. I am talking arithmetic, not even calculus.
"Let's apply GPT-4 to SVB's financial statements so we can see if they're really that bad."
*computer explodes*
"Ah, I see it worked."
Have you tried to add the profit&loss statement to the balance sheet input. Would that add more insight for gpt4.
And in the svb case what would happen if you add specific duration mismatch between asset and liabilities? Very busy interesting.
I didn't try that. Did you try?
I tried with GTP 3.5. Interesting result. Yes, GTP understooth the Income Statement provided and the most revelant lines (revenue, direct cost, overhaeds and EBITDA) and GTP suggested the direct costs were too high compared with revenue. :) Imteresting.. It was true.
But, I tried the same with other companies, and I did not get the same detailed analysis. Maybe with GTP4....
Interesting, promissing, two simple/basc compared with a real human (experienced financial analyst). But helpful when someone do not have any knlewlege about financial analysis, I would say. Maybe helpful to analyze hundersds and hundreds of companies very quikly.
hi, I am ACCA from Pakistan. I can do analysis of financial statements. I would say that GPT4 is good, it is even better than many people I know. Though it is not domain expert (I would not take it's guidance against an ACCA member who has worked in industry for over 5 years) I would still ask it to look out for dangers while making my own assessment.
Also, your assessment of poor financial strength from a very old balance sheet is incorrect, GPT4 managed to give better assessment. The reason bank is in finical distress can't be assessed from this very OLD BS i.e. there might have been many other things during that time that lead to this situation.
Covered langchain yet? An indepth look of how it works would be nice.
I think more quantitive indicators like calculation of solvency ratios etc could have been more useful
Thanks for the suggestion
Keep suggesting and make it better then we won't require equity analysts
it's done a good job despite the assessment that the bank is in good health. because it was indeed in good health, and then there came the anti-crypto politicians who ruin the banks crypto environment, and then there was a bank run, and finally, the bank went under.
Please make video about llama index
I have one before - ruclips.net/video/9TxEQQyv9cE/видео.html
But not an extensive one
@@1littlecoder thank you bro, wanna finetune docs
Financial analysis involves substantially more than a simple balance sheet. The sort of problems which led to SVB failure is not in explicit in the balance sheet. You need to understand the matching of assets and liabilities with respect to duration and timing of cash flows. May it is in the notes to the financial statements, may only management and the Fed has that information…. In short you are feeding insufucient information to GPT4 and expecting it to find things beyond what is possible with what you gave it…
Hi there, great feedback. What else can I feed chatgpt for a more accurate result?
Please use ur screen away from the chat where u r typing
While an interesting exercise, for the most part I don't see GPT-4 adding any value here. For example, it's recommendations are pretty standard actions any competent finance student would say regarding methods of mitigating risk. Impact factors are also straightforward calculations based on the balance sheet. What I find somewhat interesting is the likelihood factor and how GPT-4 is coming up with that number from its training data.
that was a terrible balance sheet analysis - it said nothing incisive. bland, obviously written by a bot with no skin in the game. didn't even compare it against other Banks. well done video info - good test.
Thank you for the feedback. any tips on how to evaluate it better?
Agreed, this kind of analysis is only good for those who are not in the financial analyst world.
I mean yeah it looks very good and comprehensive at first. But still far away from what a proper financial analyst could produce.
aaaaansk 😫😫
I really don't understand these prompts that use the word "please". It adds absolutely nothing useful...
not useful for prompt but useful to your traits.
Useful for AGI not targeting you in the future because you were pleasant to it as a baby.
Sorry its not detailed. Its superficery. What you can check is their Debt, leverage , gear ratios, Dupont analysis, RoA, vs RoCe, etc, these all things determine whether balance sheet going in right direction or not and comlany makeing enough cash to pay its payables. Thanks
anything specific I should've checked?