I ran a simple Asset Class Allocation backtest on Chat GPT. Results are interesting

General
1

Asset Class Allocation Backtest (20th May 2019 – 19th April 2025) - Annually Rebalanced. I chose 20th May 2019 as a starting point as it was just before the election results of the new Govt.

XIRR vs. Estimated Max Drawdown

  • 80/10/10 (Equity/Debt/Gold) → 11.86% XIRR, 40.1% Drawdown
  • 60/20/20 → 11.70% XIRR, 30.7% Drawdown
  • 55/25/20 → 11.65% XIRR, 28.3% Drawdown
  • 40/30/30 → 11.54% XIRR, 22.1% Drawdown
  • 35/50/15 → 11.43% XIRR, 18.8% Drawdown

Equity - Nifty BEES
Debt - Bharat Bond ETF
Gold - Gold BEES

Returns for either allocation aren’t far away from each other, but drawdowns are significantly different.

2

I ran a simple Asset Class Allocation backtest on Chat GPT.

What do you mean by “ran on ChatGPT” ? :thinking:

(Hopefully, not just prompting it to “Run a backtest …” and
accepting the plausible but not guaranteed to be factually accurate output
without fact-checking it against actual data-sources.)

3

I simply prompted GPT to do the math with various allocation. The Instruments I have selected have been around for more than 15 years.

4

Hmmm… please do cross-verify the reported results.
LLMs often fail at basic math and logic, let alone calculate XIRR and dropdowns.

Source: A Categorical Archive of ChatGPT Failures.
( The above paper is a bit dated,
but, over the past year or so, nothing relevant to this behaviour
has fundamentally changed in the underlying LLM architecture. )

1 Like
5

Thanks for sharing the paper. I wouldn’t put too much time into cross verifying as the results are fairly in line with the scheme presentations of most multi asset mutual funds. My XIRR numbers are a bit on the lower side as I have chosen a certain market peak as starting point. The XIRR changes to some degree when you chose major market bottoms as starting point (which is rarely the case for most investors).

6

OK, ignoring whether the numbers are accurate,
i.e assuming these are the actual numbers,
What do they tell us?

Couple of things this tells us is -

During this specific period of 6 years,

a. NiftyBEES had the most drawdown, compared to BharatBondETF and GoldBEES.

b. For this specific set of start-date and end-date,
these 3 instruments currently have had very similar returns.

Q. Considering all other various periods of 6 years
(i.e. repeating what was done once in the above scenario, for a variety of start and end dates),
how often is the above a NOT true? (or is it almost always true?)

7

Compare the manual allocation performance with asset allocator FOFs such as this:

8

The ChatGPT generated XIRR and drawdowns don’t look accurate to me.

You can use the below tool to backtest various portfolio asset allocations

9

It’s obvious increasing debt allocation will reduce drawdown

10

20/10/70

11

In couple of weeks this is 2nd Trading QnA post which talks about asset allocation. Its pleasing to see investors are understanding importance of proper allocation. :+1:

Its said that in 15 years one can see two cycles of up and down turn of market . May be its worthwhile to run the test from the point when reliable data (or its proxy) is available cover many more scenarios

I assumed its one year rolling returns for all possible start and end dates fin those 6 years.
Its good Question. @Anubhav_Sarkar if you have bandwidth you can re-write prompt for longer duration and specifically mentioning one year rolling returns.

yah… Agree with you.

Umm… its more than that.
Depending on how un-correlated behavior of these asset at that point of time periodic re-balancing might stabilize gains also. That’s why annual re-balance mentioned by OP plays important role keeping XIRR more or less unchanged even if equity portion is reduced from 80 to 35

1 Like