Well, my last column provoked a reaction. Not in the same league as the report by Citrini Research I’m glad to say, which predicted a world of white-collar zombies eating babies instead of Deliveroo. I merely said that AI was superior to financial advisers.
This week, I’d planned some additional prompts in my quest for an optimal portfolio. ChatGPT had made an impressive start, but I had a tonne of follow-up questions. Before that, however, I have to address the hundreds of emails I received.
Five issues popped up the most. First, I was surprised by how many readers admitted to using large language models already for investment advice. If that’s the case, why read my nonsense each week, I wondered?
Then there was the predictable response from financial advisers. Artificial intelligence will never replace the “human touch”. It cannot do the necessary “beneath the hood” due diligence. How would the models react “in a crisis?”
Advisers would say that, wouldn’t they? And if I were them, I would emphasise the value of long boozy lunches that no AI models can replicate. Let’s see what happens. More interesting were the next three concerns loads of people had.
Many of you pointed out that it’s one thing having an ex-investment professional inputting the prompts. But how does your average retail punter know what to ask? My answer to this is that ChatGPT would have come up with much smarter questions than me.
In practice, even the least sophisticated saver can type: “Hey look, I haven’t a clue about finance, but I want to go on cruises and stuff until I die. Can you help?” At which point an AI agent takes over — analysing bank accounts, risk tolerance via questionnaires, tax status, family situation, health records and so on.
The right questions would be asked. Far smarter than yours. Far smarter than mine. But there’s a catch — as loads of you were quick to identify. How does AI get around the fact that regulators often bar the purchase of many investment products without a green light from human advisers?
Indeed, parts of investing are a closed shop — and in my opinion this should change now that AI leaves everyone else for dust. Still, using ChatGPT, Claude or Gemini to build a core portfolio of equities and bonds is a huge leap for retail investors. Let’s not be too greedy yet.
This brings me to the most intriguing question hundreds of you emailed to ask: if I give different LLMs the same prompt, do they give the same advice? There are many ways to skin an investment cat, but who would trust AI with their money if the answers are wildly different?
Before I could even fire up my computer, Dr Bruce Lloyd, emeritus professor of strategic management at London South Bank University, had run my initial prompt across ChatGPT, Claude, Gemini, Grok, Copilot, Perplexity, Qwen and DeepSeek.
The results? Like their human forebears, each model had its own approach to the construction of portfolios. Copilot and Gemini, for example, were more quantitative and risk obsessed. Grok and ChatGPT seemed to prefer straightforward, low-cost implementation.
Qwen, meanwhile, was very forex aware. Perplexity focused on hedging and made sure to look at everything through a UK lens, which I liked. DeepSeek is a massive believer in diversification and classic portfolio theory. Fair enough to each.
More granularly, Gemini is big into style tilts and the use of factors such as small-caps and growth. There is also a big difference across the board about the role of alternatives and real assets — Copilot reckons I should have a fifth of my portfolio in them. At the other end of the spectrum, Perplexity recommends a 10 per cent cash holding for “psychological comfort”.
Overall, though, each of the models ended up at about 60 to 65 per cent in equities, with Copilot and Perplexity being outliers at 55 per cent. DeepSeek and Claude were happy to go into greater detail on regional splits, while Gemini and Copilot emphasised holding quality and low-volatility shares.
On the fixed income side of things, Grok and Qwen were as high as a 35 per cent weighting, with a chunky 15 per cent in corporate credit. Claude said I should only have 20 per cent in traditional fixed-income securities. The former pair also agreed that hedging back into pounds is a must.
These aren’t massive differences in asset allocation. Nor are there any moonshots (60 per cent in emerging market debt!) or ridiculous calls such as the selling out of all my equity funds last year. Everything made sense and was logically supported.
It is comforting that all the models knew the basics. That diversification is key to maximising risk-adjusted returns. The power of disciplined rebalancing. Tax efficiency. Also, what I wrote about not long ago: that timing your entry point is important — avoiding dollar cost ravaging.
God bless him, Dr Lloyd then asked ChatGPT to build a combined “best of all worlds” portfolio blending the output of all the models — from the pointy-headed rigour of Copilot and Gemini to the practical guidance of ChatGPT and Grok given that I live here in Britain.
This ain’t investment advice, but here is the answer: Vanguard FTSE All-World ETF (45 per cent); Vanguard FTSE UK All Share ETF (10 per cent); iShares MSCI World Quality Factor (10 per cent); iShares Core UK Gilts UCITS ETF plus iShares Global Aggregate Bond GBP-Hedged (25 per cent); iShares UK Property ETF plus HICL Infrastructure (5 per cent); iShares Physical Gold ETC (5 per cent); cash (5 per cent).
Would I ever implement this portfolio? Maybe. But thinking about it revealed to me a flaw in AI-driven investing. There’s no one to blame when it goes wrong.
The author is a former portfolio manager. Email: stuart.kirk@ft.com