FREE NEWSLETTER
I’ve been dabbling in AI. Began using precursor “Expert Systems” about 20 years ago, but the new apps are more generalized and interesting. I’m aware of the limitations and anyone who wants to use something like Gemini or ChatGPT should also be aware. They can (and do) generate false information with apparent confidence. This can deceive users. Such disinformation has been given the name “hallucination” or “confabulation” by AI experts. Interesting names for inaccuracy.
However, using precise prompts seems to improve the response.
I’ve been running tests on Gemini using a range of prompts to build an in-retirement portfolio. I call these “tests” because I have my answer to compare this AI expert to. My early explorations indicate this might be a useful tool to add to my other modelling methods such as Monte-Carlo simulations. I’m also exploring approaches to a more robust withdrawal strategy. All of this is to aid my younger spouse.
If anyone is interested in Gemini’s complete response, simply copy and paste my prompts into a query. However, AI being what it is, I expect every response will be slightly different.
Adding a few terms to the prompts does generate differing results. For example, I asked Gemini to:
“Provide a model of an in-retirement portfolio. The portfolio is $1,500,000. Annual withdrawal is 5%. The portfolio is to have a duration of 25 years. Provide an optimal mix of stocks and cash or bonds. Stocks may include ETFs, individual stocks in the US and global. Minimize risk.”
I used 5% withdrawal as a means to stress the response. I received a 1,000 word response. A statement of goals and assumptions (inflation, returns and withdrawal rate) were made. Then Gemini went on to provide an “Optimal Mix of Stocks, Bonds, and Cash (Risk-Minimized Focus)”, a “Detailed Asset Allocation”, “Withdrawal Strategy” and “Management & Rebalancing”. The portfolio was 50/50 bonds-cash/equities. It used specific ETFs in its examples.
Gemini summed its response this way: “This model provides a robust framework for a risk-minimized, in-retirement portfolio designed for a 25-year duration with a 5% annual withdrawal. The emphasis on high-quality bonds and diversified equities aims to generate income, preserve capital, and offer some inflation protection, all while prioritizing risk management.”
Adding Dividends. I then modified the query: ““Provide a model of an in-retirement portfolio. The initial value is $1,500,000. The initial withdrawal rate will be 5%. The duration is 25 years. Consider inflation and the possible returns from US and global stocks, as well as cash and bonds. Should dividend paying stocks be included and if so, what percentage of equities?”
I received a more detailed 1,500 word response. This portfolio was 60% bonds-cash and 40% equities. It included a list of example ETFs for stocks and bonds. The allocation percentages of each were provided. A Bucket withdrawal approach was outlined.
Gemini summed it this way: “This comprehensive portfolio model balances income generation, inflation protection, and growth, aiming to provide a sustainable retirement income stream for 25 years while effectively managing risk and leveraging the benefits of dividend-paying stocks.”
Following Mark’s thread yesterday I tried it out with a hypothetical with numbers and a mixture of sources to draw from to optimise income over a hypothetical 35 year retirement.
Gemini didn’t do badly – basically regurgitating sound traditional advice on draw from tax-free and after tax sources first and advised on a conservative SWR. I prompted it to go a bit deeper on strategies re accepting more tax at low rates earlier and refilling our Roth equivalent and it struggled a bit probably because it couldn’t really anticipate inflation and how this might outpace tax rate bands.
With enough prompts and rework I can imagine it producing a passable simple strategy that you could then test with modelling.
I posted an article yesterday on some of the possible dangers to consider with AI financial advice.
Gemini states “Gemini can make mistakes, so double-check it.”
I view this as a source of ideas but I do think I can refine the output for specific purposes. I will not be using the answers (output) to build my portfolio.
One of the challenges is that Gemini uses web based resources. For example, I used a model portfolio in a query about optimizing portfolio withdrawals in retirement using a bucket strategy and what to do about dividends and capital gains: move to the cash bucket or reinvest, and under what circumstances and the steps to replenish? Etc. The response to a basic query was about 1,900 words.
Using the “Deep Research” tool in Gemini I received a 4000+ word report after it scoured about 30 websites. Now, using data from other sites does mean it is possible some “bad” information will be provided as copied from these sites. Gemini does provide a list of websites used, as well as those it “chose” not to use. (The research took more than 12 minutes. I wonder how many kW were used?)
I am using different AI tools to get ideas, although some of the coding I’ve been provided is brilliant in that it took little effort on my part to get the code generated. Coders beware!.