Imagine standing in front of your boss and recommending the purchase of a $5 million machine learning tool for fighting fraud that your financial institution would never be able to use. Sounds ludicrous, doesn’t it? And yet, that’s akin to what Netflix did.

The “Netflix Prize” AI platform build project

In 2006, Netflix announced the Netflix Prize, a machine learning and data mining competition. Netflix offered $1 million to any person or team who improved the accuracy of its movie rating prediction machine.

It took three years for a developer team to build a machine-learning algorithm that met the requirements to win the contest and increased Netflix’s recommendation engine by 10 percent. But by that time, Netflix couldn’t use the $1 million dollar engine they had bought because the engineering costs involved with implementing the new algorithm were prohibitive. Also, in the span of those three years, the industry had shifted dramatically enough that the algorithm was already outdated.

If the Netflix Prize story was an Aesop fable, the lesson would be clear: there are more things to consider when looking at machine learning than just KPIs.

You might be thinking, “We’re better than Netflix. We have a better team. We can build a machine learning platform.” Maybe you are, and maybe you can. But I’d remind you that Netflix disrupted the video rental industry, pivoted to original content when major Hollywood studios denied them titles, won an Oscar six years after that move, and grew their audience by 70 million subscribers in seven years. They’re not stupid.

To get machine learning right, you have to be a machine learning company

Machine Learning is challenging for most organizations. This technology is a far cry from standard computer programming. Basically, to build a machine learning platform for risk, your organization needs to be, at its core, an AI company. This means you have deep expertise in artificial intelligence, commit the resources to build out a team of data scientists (in and of itself a challenging task as there is a worldwide shortage of these skills), build and maintain a massive infrastructure — along with the team to maintain and scale it — and assure regulatory compliance across the platform.

That’s a tall order, and it almost never works when the organization is a financial institution that happens to have a machine learning team.

Yes, a small number of organizations have been able to do it. But overwhelmingly, organizations fail at building their own platform for risk. We have organizations come to us after they’ve sunk about eighteen months of time and millions of dollars trying to build their own machine learning platform.

As Andrew Ng, co-founder of Google Brain so perfectly expressed, “Any company + deep learning ≠ an AI company.”

Why it is so challenging to make machine learning work in risk and in banking

The road to production is not a linear, easy process. It’s a four-phased approach that includes preparation, implementation, validation, and production. Each phase is complex in and of itself.

For example, in the preparation phase, you must pay special attention to the proper building of your infrastructure. You need to ensure you have enough storage and sufficient processing power. And, when you’re looking at doing this for a leading bank, payment company, or any other major business, you need it to be scalable and flexible to meet future evolutions of your business. The bottom line for this phase is that you can’t afford to underestimate the time and resources data preparation takes.

And that’s just one phase.

What does all of this tell us? You need a machine learning platform that is optimized for risk and the financial ecosystem that addresses all of the above points.

How long does it take to build a machine learning platform for risk?

You should expect it to take about five years to build your own machine learning platform. Now, ask yourself, how will your business’s use cases evolve in the next five years? What about the industry as a whole? You’re going to have to be able to predict these things in order to build a platform that will be useful five years down the road.
Make sure to build a useable machine learning platform for risk

And that brings up the next point: you don’t just have to build a machine learning platform, you have to build a usable machine learning platform for risk. As Margarida Ruela, Director of Product, AI, at Feedzai cautions young data scientists in workshops, “Buildable does not equal usable, which does not equal really usable.”

In order to be “really usable” in the financial industry, you must ensure you have the right data, the data is pristine, and that decisions made based on your machine learning models are explainable to regulators. There is no wiggle room here. Your models must generate trustworthy and accurate explanations.

Rather than build vs. buy, think machine learning partner

Still, you have legitimate reasons for wanting to build your own machine learning platform for risk. And, they’re good ones. But instead of risking your personal reputation on building a machine learning project, why not use those reasons to vet your machine learning risk platform partner?

The right partner can actually teach you how to build and manage your own platform. But how do you choose the right machine learning platform partner?

Choose a machine learning partner that provides you with:

1. Control
You want control over the data schema, different integrations, all the layouts, formats, which models to use, and the cadence that you can update profiles. Make sure you are able to self-manage your dashboards and reporting so you can see the right metrics and KPIs relevant to fighting fraud at your organization.

2. Flexibility
Make sure your partner has the ability to take on new use cases so you can deal with tomorrow’s problems and opportunities today. Also, look for a partner that allows you to run multiple models. Fraudsters behave differently depending on where they are and what they’re doing. Your machine learning risk platform should be as flexible as the criminals are.

3. Freedom from vendor lock
A lot of risk platforms require financial institutions to use specific vendors, and that can be a hassle.

4. Production speed
Partnering with experts in machine learning platforms for risk allows you to condense five years of engineering work — the realistic time frame for building your own ML risk platform — into five months.

5. Process
Building a machine learning platform is an iterative process. Planning and timelines won’t mimic other projects your team has managed. Your platform partner should help you develop processes and documentation so that your system works regardless of employee turnover, organizational pivots, or other unforeseen changes.

Machine learning increases revenue, enhances customer experience, and reduces operational costs. Partnering with the right machine learning experts today can help evolve your organization into an AI powerhouse tomorrow.