How Robo Advice works

To trust robo advice, you have to understand how it works. It’s not as mysterious as it sounds.

ROBO ADVICE sounds like the next big thing. Journalists are writing about it on a daily basis and big banks here and around the world say they are working on ways to automate wealth management using robo advice as a lever.

In science fiction, however, we know that robots are generally not to be trusted. Whoever came up with the term “robo advice” should have tried a bit harder. If someone could explain how the algorithm embedded in this metaphorical “robo” works, it might be easier for any doubters out there to get with the program.

That could be like asking a mathematician to turn numbers into words. In fact, that’s exactly what it is. So here goes.


The mathematical formula robo advice is built on is nothing new. It was developed in the 1950s in the US and was quickly adopted by portfolio managers everywhere. Harry Markowitz’s “mean-variance” model is an important chapter of Modern Portfolio Theory, a curriculum staple at any university with a business faculty. The mean-variance algorithm relies on some basic statistical measures known to all high school students: mean (average), variance (or standard deviation) and covariance (a measure for the relative variance between two variables).


The model isn’t some magical pudding that just produces perfect portfolios. It’s an equation. At the end, it will find the weightings that will deliver a set return for minimal volatility. The weightings are the percentages of your starting capital to invest in each asset.

Of course, the model has no idea what will happen to values in the future. There is no chance of it producing the best portfolio, but it will design one that is the best fit for the past. What’s the good of that, you may ask? Well, past data is known — no-one can argue about it — and if the data for the mean, variance and covariance inputs takes in a long period, then the belief is that it’s OK to make a few assumptions. If the stats are taken from the previous 30 years, who’s to say they’re not good estimates for the next one or two years?


The algorithm at the core of robo advice is designed to produce the portfolio that will deliver a fixed return with as little upward or downward variance as possible. It’s for that reason the user is asked to take a risk profile test. It’s possible to deduce a required return from an investor’s risk tolerance level because of the existence of the efficient frontier. If investors want higher returns, they must accept greater variance of returns. The efficient frontier plots this relationship.  

The robot solves for the return number input by the user, but the user must accept there is no guarantee of achieving that return.


Once the asset classes are chosen and the stats are collected, the algorithm can be set to work to find portfolio weightings that will match the expected return and minimise risk. If you input a range of expected returns, the robot will produce a bunch of different portfolios. The expected risk (volatility) for each portfolio will be different. As the expected return increases, so too does the risk. When the two are plotted, we get the efficient frontier.

Remember, what comes out is reliant on what you put in at the start. If the assets are changed, or if the period over which the data is collected is altered, the output will be different. Users of robo advice can expect providers to produce portfolios that include all the usual asset classes — Australian and international equities, listed property, fixed income and cash — but they might want to ask over what period the data is collected.


The model looks at past returns, the variance of past returns (standard deviation) and the covariance between assets. It will be common to see assets with lower expected returns also exhibiting lower price variance. The cash rate, for example, doesn’t move around much. Bonds will generally be next on the risk-return scale, with higher expectations. Listed property might be next, then shares.

Where there are deep, liquid markets — shares, for example — the data can be trusted. But it’s possible for robo advice to be hi-jacked, and consumers need to be aware of this. If an illiquid asset class is included, it will be much harder to trust the data. An example is property. Listed property in the form of real estate investment trusts are priced constantly, with units trading on the exchange, but a robo advice product which claimed to include residential property should make you stop and wonder. If the data is collected every quarter or so, the statistic for price variance could turn out to be very low. The algorithm wouldn’t ask questions, of course, and as a result the portfolio it produces might be heavily skewed towards the asset.

You have to trust the data before you can trust the algorithm. If junk goes in, junk comes out.


Robo advice for the masses has been enabled by technology. Students of portfolio theory might be tasked with replicating the algorithm using Excel, for example, but consumers are being asked to trust providers accessed via the internet.

The number of calculations going on in the background can be vast, so it suits providers to limit the variables fed into the algorithm. Instead of crunching the top 200 Australian stocks and top 500 US stocks to build a matrix with 246,050 pieces of data required for the algorithm to work, for example, a robo adviser might substitute two ETFs for the S&P/ASX200 and S&P500 indices, which relies on six stats.

Some robo advice models offer a limited range of investments, as few as five in some cases. The work done by the robot will be pretty simple, following the mean-variance algorithm, but some or all of the five investments may be ETFs. In that case, the portfolio could include thousands of stocks, but the algorithm has parsed a manageable 20 pieces of data.


The algorithm at the core of robo advice was designed to use past data, but professional investment managers who use it as a tool for portfolio construction can easily tweak the numbers to reflect their expectations for returns and volatility of returns.

Providers of robo advice might not be all that good at communicating to consumers how their expectations for the various investment classes affects the portfolios they produce, but investors should know there’s an element of subjectivity. The mathematical foundation of robo advice is well-known, but the value of the investment portfolios it produces relies on the data fed into it. A provider who has “added value” to a robo advice service may have done so by changing the returns and variance data to match its expectations for the future. Consumers have to be comfortable with that.

Providers also have the freedom to pick and choose the investment variables, so a branded robo advisor might be expected to turn out portfolios of branded investments. If you trust the brand, that’s fine. But there is a danger some providers who offer robo advice as a low-cost objective investment solution may be taking advantage of consumers’ trust.

What you get out is only as good as what you put in.

Related Articles