It was back in 2014 when researchers at DeepMind directed their nascent artificial intelligence application to the game Breakout. Instead of programming DeepMind on how to play the game, the researchers programmed DeepMind to learn about learning to play the game. That is a meta level that isn’t normally programmed but the result of that effort, combined with many others, was in the back of our mind when we turned some artificial intelligent tools towards customer data. Similar to the 1980’s arcade game, the result was surprising and lays the construct of how banker and machine will collaborate in the future.
Getting The Last Brick
The technology conference was called The First Day of Tomorrow and DeepMind (subsequently bought by Google) demonstrated their creation. The program took control of the video game’s paddle and randomly shifted it around hoping that it would connect to the square ball on the screen. If the ball hit the paddle, it would then bounce up to a brick above, and the brick would disappear. Each brick represented a point, and the goal was to try to eliminate all the bricks and make it to a faster-moving level. Of course, hope and randomness are human traits and were attributes given by the audience in order to make sense of what this alien intelligence was doing. DeepMind was beyond that.
DeepMind tested each position on the screen and within a couple of minutes, hit its first ball. The crowd went crazy.
After 30 minutes, around its 200th game, DeepMind missed the ball about three out of four times. This was a learning curve that was much shallower than your average teenager trying out Pokemon Go for the first time. However, after an hour of playing, DeepMind was just getting going and then never missed the ball. It was during that hour, that DeepMind was also processing in parallel the optimized way to take out each brick. While most players start by taking out layer after horizontal layer, DeepMind would work the sides of the wall focusing its efforts in a vertical breakthrough. By the end of hour three, most all of human knowledge as it pertains to Breakout was now learned by DeepMind. The program was now teaching humans how to play as onlookers learned that the left side of the wall provides weaker than the right due to a programming flaw. DeepMind wasn’t done.
The program then took its Breakout knowledge and applied it to the other 49 Atari video games quickly beating humans in more than half of them. As a learning process, hope and randomness were both vanquished.
The Deeper Meaning Of DeepMind
Now playing the perfect Breakout game may not seem like an achievement, especially when you consider IBM’s Deep Blue past success in chess but it is. Where Deep Blue achieved its victory over Grandmaster Garry Kasparov by computational brute strength, DeepMind mastered Atari games by intellectual elegance. What it learned from Breakout, it applied to Fishing Derby and then to Centipede. Where Deep Blue was specialized, DeepMind handled a variety of games. It is this varietal problem solving that is more akin to human intelligence.
Ironically, it was Ms. Pac Man and Gravitar, two games easily mastered, that DeepMind struggled with due to the sheer number of potential outcomes that eclipsed the program’s historic learning horizon. However, in Spring of this year, the next version of DeepMind, Alpha Go, combined DeepMind’s learning ability with Deep Blue’s computational engine and time horizon and soundly defeated Lee Sedol one of the best players in the world at one of mankind’s hardest games – Go.
Artificial intelligence, like the game changers before it - mechanical leverage and electricity – will be applied to one endeavor after another. Already surpassing static algorithms, reinforced machine learning (similar to DeepMind) and neural networks are already protecting banks when it comes to fraud and cybersecurity protection.
After testing artificially intelligent “robo advisors” for wealth management, we were impressed not only for the sum of knowledge contained in these programs, the ability for this programs to learn current market conditions, but to learn about our evolving goals. We wanted to take these same machine learning techniques and apply them to the problem of how we can make our customer smarter.
First Stop – Cash Management
We ultimately want to get to using artificial intelligence for commercial credit and customer management. However, lacking a robust enough data set, we wanted to start small. Banks can purchase artificial intelligence on a pay per use basis akin to warehouse space, electricity or digital storage. While banks can purchase standalone programs, they can also subscribe to a “platform as a service” vendor such as Amazon, Microsoft to Google. At $100 to $1,000 per month (based on transactions analyzed and computational time), these applications are relatively inexpensive and moderately easy to use requiring little programming knowledge.
Loading in several years’ worth of monthly financial statements yielded interesting results. Largely, short-term assets and liabilities are not being actively managed for the average small business. By applying machine learning to cash inflows and outflows, we learned that a business could increase its cash balances by an average of 15% and probably closer to 30%. In a low-interest rate environment, this may not much matter, but for a growing business that investment alternatives include having enough excess cash to invest in more revenue-producing assets like machines, human capital and real estate, those excess cash balances can save a business from raising more expensive equity capital.
The Biggest Lesson – Cash Flow At A Higher Level
By far the largest lesson learned in analyzing the output of artificially intelligent data analysis was the true cost of payroll. Most businesses have never questioned processing payroll on the first and fifteenth of each month. This creates huge artificial spikes in cash outflows for most businesses and by smoothing payroll out, an intelligent cash management applications can better match sales, receivables and payables to create a more stable cash flow stream.
For example, the application starts to learn cash collection cycle times for each sale so it can recommend a discount within a pricing strategy to increase cash. For businesses that are largely labor based, these discounts are minimal, as by paying certain people at different times throughout the month, discounts can be minimized except when certain situations occur. When a lump sum cash payment is required, maybe a single day discount can be applied to increase projected cash balances in order to make certain lump sum payroll days.
Of course, there is more than just payroll to consider (although that was the dominate expense for a majority of business we analyzed). Similar to discounting the sales side, receivables can also be dynamically discounted so that collections can occur at faster than the maximum terms might allow.
For different seasons, different months and different goals, what is required is a dynamic system that continues to learn. Our analysis was merely applying basic machine learning to a time series of historic data to answer the question what should the business have done. The next step is to attempt to learn the correlative factors of each business and then attempt to project cash flow needs. Once projected, artificial intelligence can help re-optimize the appropriate discounts and premiums to solve for maximizing free cash flow to the extent it can limit credit and capital.
Above is an infographic of four things that consistently came out in the data we analyzed. Of course, our machine learning tools were basic, the time series limited (two years) and the number of businesses (ten) in no way is a representative sample, but the output was conclusive enough to prove the benefits of an application of artificial intelligence as applied to cash management.
As can be seen, managing so many variables on a real-time basis is an effort perfectly designed for machine learning. Not only is the cash flow fingerprint different for every business, but new factor relationships constantly change. The influence of gasoline purchasing for product delivery may be a dominant factor influencing cash flow one month, but as we saw in our data, as the cost of energy changed, the importance was decreased over time. The nodes within the machine learning algorithm then were rearranged so that the cost of materials were more heavily weighted.
The Rise Of Machines Plus Bankers
Just like this analysis made us smarter in looking at the cycle times of cash flows, we envision a day that an application can do this with clear recommendations. This is an effort that is currently overlooked by all but the most diligent bankers and thus represents a clear path where artificial intelligence will require a new position within banking to interpret the output and to advise the business on both cash flow and capital.
Banks are in an ideal position to execute on this business model, and we are just years away from this becoming a reality for the benefit of all small and mid-sized businesses. Banks with this technology will have a clear advantage over banks that don’t and will be able to more easily garner the ever so valuable C&I business with its large cash balances and ability to generate fees from treasury services.
Just as DeepMind learned to master the learning of arcade games, DeepMind and similar applications can help bankers master all aspects of a complicated business. Artificial intelligence will soon be the single biggest game changer banks have ever seen. It will reshape all facets of the bank business model from customer advisory, customer management, investments, credit decisions, credit management, pricing and strategy. It will touch every line of business and both bankers and customers will be better off because of it.
Submitted by Chris Nichols on July 28, 2016