What would happen if you fed all sorts of financial, economic and corporate data into a state-of-the-art neural network and used it to forecast returns for the US stock market over the next decade?
What kind of magical revelations would this field of artificial intelligence — where layers of neurons mimicking the structure of the human brain spit out non-linear predictions — yield up to inquiring financial researchers?
Hold on, as this may blow your mind: The S&P 500 is likely to return 8 per cent a year over the next 10 years (pretty close to its long-term average) and stocks will outperform bonds over that time.
From JPMorgan’s long-term strategy group led by Jan Loeys:
– We construct a neural network model based on fundamentals to forecast 10-year out returns on the S&P 500.
– The model makes no assumptions about the future and is based solely upon currently observable fundamentals. This produces more objective forecasts and allows us to estimate the risk around forecasts based on historical model out-of-sample performance.
– Fundamentals include different measures of the internal rate of return on the index, macroeconomic factors, and lagged returns. Macroeconomic factors include household asset allocations, real yields, and recent economic growth.
– The model exhibits strong predictive power over 10-year returns, with relatively small forecasting errors and approximately zero statistical bias, implying the predictions were neither too optimistic nor too pessimistic on average.
– The model forecasts an SPX return of 8.0% pa in the coming 10 years. This forecast comes with a one-sigma risk of ~1.5%, implying a 2/3 probability of a return between 6.5% and 9.5% pa. This risk measure is based on the out-of-sample forecast errors of our model and is, thus, not to be confused with within-sample errors.
– This conditional risk around the 10-year compound return is much lower than what would be implied by the annual return volatility of 17.4%, and the historic volatility of 10-year returns.
– Based on the current 4.8% yield on the US Aggregate Bond market, which is the best estimate for its 10-year return, equities are very likely to outperform bonds in the coming decade from today’s entry point, producing comparable returns even with a two-sigma underperformance of equities relative to our forecast.
– Jointly, these equity and bond return forecasts imply a decade-ahead return of 6.7% pa on a 60-40 strategic allocation in US equities and bonds. This forecast comes with a one-sigma risk of ~1%, implying around a 2/3 probability of a return between 5.7% and 7.7% pa.
This reinforces FT Alphaville’s view that a lot of admittedly interesting work around using various AI approaches to tackle big economic and financial questions just tends to yield “well, doh” answers — such as the Bank of England discovering that credit booms and inverted yield curves tend to augur financial crises.
To be fair, JPMorgan’s team of Alexander Wise and Loeys did test their neural network on quasi out-of-sample data, with decent results. It had used data from 1952 onwards to train it, and then began forecasting results 30 years later — giving it kinda live results from the early 1990s onwards.
The model’s predictive powers improved over time, JPMorgan estimates. Here’s what it looked like in chart form:
But applying AI to bigger broad questions like this tends to be the equivalent to using a Rube Goldberg machine for everyday tasks: cool but overly elaborate, and the outcome often feels like a damp squib.
That doesn’t make the exercises completely useless though. Sometimes you can discovered interesting things by accident, even if the main conclusion is painfully obvious from the outset. And next time perhaps the question is more fine-tuned and suitable for the tools used.