The challenges of Robo-Advice


This is a guest blog from Martijn Vos of Ortec Finance. I know Ortec well- they are great financial modellers, I don’t know Martijn but he’s obviously been reading my stuff and as I like his article very much, I’ve asked to include his blog on mine!

In a previous post we discussed how robo advice can help with financial decision making. Making the right investment decisions is complicated. Ortec Finance enables people to manage this complexity.

We do so by providing them with the appropriate tools and advice, so they better understand the impact of their decisions and adjust them or their goals when needed. Robo advice is such a tool.

Of course, developing such a tool is not easy. Robo advice is based on input parameters, uses models and algorithms, and creates output that must be interpreted. There are challenges in each of these steps. In this post we focus on the role of models and algorithms in robo advice and we assume that the robot will restrict itself to an automated investment decision advice. In a future post we will address automated holistic. In addition to a recommendation regarding the allocation decision, the robot typically includes information on how likely it is that the underlying goals can be reached.

In order to create an investment advice including its impact on goals, the robot will use models. For the problem at hand Asset Liability Management models are the standard. In a recent overview of the ALM models that are available, the authors consider individual ALM as a synonym for ‘robo advice’ or ‘custom defined contribution’. They label these models as complex, requiring at least 200 input variables.[1].

Models and algorithms requiring assumptions

We certainly do not deny these models are complex, but so is the problem. In the context of pension design Keith Ambachtsheer once quoted Einstein saying that models should be as simple as possible for the problem at hand, but not any simpler[2]. This is one of the key dilemmas in robo advice. Models should be applied with care (e.g. parameter uncertainty, black swans, costs etc. can all have a big impact on the outcome), they should be fitted to the problem at hand, and communicating their results should be simplified and tailored to the situation as well.

What does a simple as possible model for robo advice entail? Well, first of all, it needs input on the future goals (of the client); as well as information on the current situation (horizon, current and future income, current assets etc.). One of the advantages of robo advice is that the advisory process becomes more consistent and objective[3]. This helps advisors with transitioning from a product-oriented to a client-oriented approach. It is also important to include fiscal regimes, state pensions, etc.? Finally, and most importantly, we have to make assumptions regarding the future state of the economy and financial markets. Scenario analysis is the tool most often used to do this in the context of robo advice.

Scenario analysis needs to take into account all elements that impact the consequences of decisions as realistic as possible. Furthermore – and this is much more challenging and open to debate – one has to construct scenarios that describe, as realistic as possible of course, what might happen to economies and financial markets in the future.

We have to mention two lessons we learned from this scenario approach. The first is that it is very important to be as objective and explicit as possible about the (scenario) assumptions one makes about the future. Scenarios should “just” be the most realistic description of the future we can make. The second lesson is to always keep monitoring the progress made towards the (end) objective. If one has used assumptions that are too optimistic, and financial markets develop in negative directions, this will show in decreasing probabilities to reach the objectives.[4]

The output of the robo advice

Once we have created the input for all of the above we can use an algorithm (the actual robot) that processes the input to output. Henry Tapper recently phrased it brilliantly: [5] he explains that an algorithm cannot lie: An algorithm will tell you what you set it to tell you with the data you fed it. It has no choice, it is entirely deterministic; and yet algorithms can help you lie. Why, because the assumptions and data underlying the algorithm can be false! All models that describe reality are, naturally, abstractions of this same reality and it is not only the assumptions but also the use and interpretation of the model outcomes that defines the quality of the (investment) decisions; see for example Boender et al (2011) for an overview[6].

So, now we have jumped the hurdles of the individual input and the input and design of the models and algorithms, the final step is to present the outcome or actual advice to the individual. In a next post we will address ‘robo’ solutions to this step as well. In the meantime, we love to hear your feedback.


[1] Agnes Joseph, Niek Paanakker, Jop Versteegt, Van collectief tot individueel ALM, PBM Dossierreeks 9, Pensioen Topics, 2016

[2] Keith Ambachtsheer, Taking the Dutch pension system to the next level: a view from the outside, Netspar Occasional Paper 2014.

[3] See e.g.

[4] More details in: Relevance of scenario models, Hens Steehouwer, May 2016


[6] (in Dutch)

About henry tapper

Founder of the Pension PlayPen,, partner of Stella, father of Olly . I am the Pension Plowman
This entry was posted in pensions. Bookmark the permalink.

Leave a Reply