5. Learning Machines: Letting a machine help

A learning machine is like a savant. It can be really good at one particular task. So building a simple machine trained to help guide my investment decisions is the path I've chosen. Learning machines are part of the broader class of products associated with artificial intelligence. InvestEngines has been designed to uncover patterns in the data that are not intuitively obvious and to use those patterns to assist in making investment decisions.

I'm not going to boil the ocean here, but its worth having a basic concept of what a learning machine is. If you want to build a machine to find a cat in a picture there are two basic ways for the machine to learn and both ways start with lots of data (for the cat example, lots of pictures of cats).

One way is to give the machine some help in the form of algorithms (like what cat eyes and ears look like) and see if the algorithms are sufficient to find cats in pictures. The other way is to build a digital equivalent of a set of neurons (you know, like in your brain) and let the machine find the connections that deliver the desired result. It is iterative and takes a lot of time to initially process. For purposes of investing, I've built very primitive and super simple versions of both. And as I've discovered, even these simple machines offer some investing insights and quantitative guidance that other methods do not.

InvestEngines 1.0 was built with an emphasis on algorithms and InvestEngines 2.0 was built with an emphasis on regression and neural weighting ideas. Taken together, they have delivered noteworthy results. InvestEngines 2.0 took advantage of lessons learned from the vast amount of data processed in the first version to refine its models. Taken together they offer some compelling insights.

For most long term investors, simple diversification is the best approach, but occasionally there are moments in time when taking action to dramatically realign asset allocation can prove advantageous. Those times can appear when major market dislocations occur, such as the tech bubble bursting and the financial crisis. Opportunities also appear when a particular asset class tends to outperform for an extended period of time. So, naturally, the first 2 questions to be answered at the moment are if any asset classes are likely to outperform on a long term basis and/or are we near a bubble bursting or a financial crisis. And if so, is it statistically worth taking any action. If the answer to both questions is negative, then broad balanced diversification is still the optimal path. If either question is answered in the positive, then there may be some compelling benefit to consider over-weighting an asset or avoiding a particular investment altogether. That is the primary output of the learning machines.

There are no guarantees and there are no free lunches. So, a word of caution about learning machines in general and InvestEngines specifically. A learning machine can do no better than the data that was used to train it. In the case of investing, we only have historical data to learn from. And as any investment professional will tell you, past results may not indicate future results. And they will also tell you, any "system" will work until it doesn't. I agree with both points. So I've built two major elements into InvestEngines 2.0 to deal with those realities: A utility function that maximizes consistency over time and a self checking fail-safe that returns to the benchmark portfolio when the learning machine fails. Only time will tell the degree to which they are effective.

With that cautionary note, we will look at more details on how InvestEngines has performed over the past 20 years and what it is presently indicating in the following articles.