What Good is a Computer?

In the course of our business, helping PMs get better results from their analysis and decisions, we often hear this:

“Oh, I have used an optimiser, it didn’t help/ didn’t work/produced nonsense numbers”

This highlights reason #2 of my list that answer this question:

Why it is hard to implement a systematic approach using mathematical tools for portfolio construction, even though the resulting risk and reward characteristics are very clearly better?

3 reasons that make it difficult to institute a systematic approach to portfolio construction.

1: Organising Data so that you can use a computer to help solve the problem

2: Asking the right question of the computer so the answer is useful

3: Visualising the output in a way that makes it clear to all stakeholders that your portfolio construction does what you want it to.

So why is it hard to ask the right question?

Computers are excellent calculation engines: they don’t make mistakes, the results are reproducible and the process by which they arrive at them is mostly (save for AI type Neural Networks etc.) well understood.

But that doesn’t mean they can be left to do it all by themselves. Quite the opposite: And the more complex the output, the more care you must take with the inputs: as the problem becomes more highly dimensioned, (i.e. the output contains more numbers that inter-relate) asking the question correctly becomes much more important.

The opposite is true of AI systems (Neural Networks or similar) that are used to classify data (‘this image is a cat or not-a-cat’, ‘that object is a car or it’s the shadow of a car’, ‘this is your face or someone else’s face’), the output is very low dimension, a binary yes or no, and the quality of the input is less important. Indeed, the purpose of the AI is to handle undefined inputs of varying quality (video, images etc) and generate a low dimension output with decent certainty.

But business problems, including Portfolio Construction, are the opposite.

For example, a logistics or information processing problem (say ERP in a manufacturing context or planning a distribution network) has a very high dimensional output. You need to get all the parts/people/goods to the right place at the right time with the right instructions/packaging/tracking number. The output is a large data set, maybe 000’s of numbers that control a factory’s day-to-day operations or perform accounting functions in a bank. ?

In this sort of problem the inputs are well known (i.e. the unit-cost or time-taken for a process), and the maths is simple (mostly linear), but any errors in the input are potentially multiplied by the maths and lead to bad outcomes.

If you have given the system the wrong unit cost for an operation, the system will find a solution that penalises (if too high) that operation, and this will throw all the other operations out of kilter too (as they all inter-depend), no matter how obviously wrong it is.

So, with these types of business calculations, getting the inputs right is essential. That’s what all the consultants do when you implement a new logistics system.

And in Portfolio Construction, the same applies.

In portfolio construction, the inputs are reasonably well understood, but they come with uncertainty that doesn’t exist in a unit-cost calculation. Analyst recommendations or conviction scores are known and, in aggregate, contain information but also have uncertainty: they are not precise.

Portfolio construction uses more complex mathematics than logistics, though the number of outputs is lower: the portfolio has maybe 30-80 interdependent numbers, (i.e. the number of assets).?There are typically around 3-5 attributes per asset (like sector/country etc.), so in total we are dealing with 100-400 data points. (plus whatever market data history you use).

The calculation may be non-linear (at Sherpa we use asymmetric risk functions to better model the real-world implications of P&L and stochastic simulations of future markets over which we find the best stable result), but it is in essence they are easy to understand: you are looking for the portfolio that best balances your convictions against the concentration of your risk.

So portfolio construction calculations straddle the 2 worlds:

You have imperfect information (it’s better than the info that AI analyses, but worse than a logistics operation), but there’s not that much of it. You have non-linear calculations, and you also have a high dimensional output with high interdependency.

This is why it’s difficult: computational solutions to this type of problem multiply the effect of input errors, whereas AI type solutions to classification problems reduce input errors.

And this is why people have such trouble using computational solutions. The plug-and-play approach: ‘here’s your portfolio-calculator, press Shift+F9’, will never give you good results.

You have to focus on the inputs, you have to understand what your convictions and constraints really mean. If you do that, then you can use a computer, and it really does work. Sherpa’s clients see returns consistently 200-400bps better p.a. from using our computational approach combined with our detailed attention to understanding the problem correctly.

At Sherpa we look at Portfolio Construction holistically: We spend a lot of time with our clients going through the inputs, the true meaning of their asset election and convictions, creating fully defined consistent constraint sets BEFORE we get anywhere near the mathematics. We recognise that this time is spent ensuring that the problem is fully scoped so that the output is sensible.

Implementing pure computational solutions doesn’t focus on getting the inputs right.

Pure human solutions recognise the lack of precision in inputs and sacrifice performance in order to deal with it.

But a properly managed computational solution bridges this divide.

Ask Sherpa how we can help you gain an extra 200-400bps pa.