Decision analytics for the perfect race

By | 6 minute read | June 22, 2016

Dave Haase

When fellow IBMer Doug Barton asked me whether we could help extreme cyclist Dave Haase win the Race Across America (RAAM), I immediately agreed. We keep saying that big data and analytics help people make better decisions, and RAAM presented a great opportunity for us to prove it.

Three questions sprang to mind at once:

  • What data can we access?
  • What insights can we gain from that data?
  • What key decisions does Dave need to make during the race?

That last question is a great place to start when undertaking any big data analytics project: What decisions do we want to improve to get the outcomes we want?

The race

What are the right questions for Dave to answer? To find out, we first had to get acquainted with RAAM. We learned much through numerous interactions with Dave; you can read a good summary in Doug Barton’s blog post “Sweat grit and analytics to get to the finish line first.”

RAAM is a bicycle race, more than 3,000 miles long, from Oceanside, California, to Annapolis, Maryland with no planned rest periods. The route is fixed, but each racer decides when, where and for how long he or she stops to rest along that route:


The decisions

After much analysis, we agreed that the most important decisions were when and where Dave should rest. He will race for eight or nine days, sleeping only two hours a day. Incredibly, he will ride the other 22 hours. If I were driving a car over the same route, I doubt I would be able to finish before Dave: I would need at least eight hours of rest each day, not including breaks for meals and other necessities.

Dave’s ability to ride virtually nonstop for days on end may seem machinelike, but he’s no machine. Dave will tire. His power and speed will decline as he races, and eventually, he will have to rest to restore his power. After resting, he will be able to ride at greater speeds once more.

So why not rest more often to maintain a high power level? Simple. When Dave rests, he is not moving. And as the minutes go by, so can other riders! So we must balance two competing goals:

  • Have Dave rest, restoring him to a higher level of power to increase his speed.
  • Keep Dave on the course as much as possible, increasing his distance covered.

Making decisions while balancing conflicting goals is a great use case for decision optimization, an analytics method focusing on computing the best options in a given situation. So we developed a decision optimization model that helps find optimal times for Dave to rest during RAAM. This model sees Dave as an engine whose power declines as he rides and increases when he rests.

Internet of Dave

To derive the decision optimization model, we need data about how Dave’s power rises and falls when he rides and rests. This is where the “Internet of Dave” comes into play. As explained in Doug’s post, Dave wears sensors supplied by Garmin and Equivital on his body when he rides. These sensors provide data about Dave’s physical condition, including his power and speed. We collected this data over months during Dave’s training rides. Then we analyzed it using tools such as Watson Analytics, yielding solid estimates about how Dave’s power evolves.


The analytics advantage

We have closed the loop, answering our three questions:

  • What data can we access? Internet of Dave sensor data, including Dave’s medical condition and the power he is able to deliver.
  • What insights can we gain from that data? How Dave’s power evolves when he is riding and when he is resting.
  • What key decisions does Dave make during the race? When to rest to help him ride at maximum speed and finish the race as quickly as possible.


Using physics to our advantage

The effects of wind are obvious to even amateur cyclists. When the wind is coming from behind (a tailwind), a rider will go faster than when riding into a headwind. The difference can be dramatic. The slope of the road is also important: Climbing, of course, is much slower than going downhill. And quite often these two factors work together.

To evaluate the importance of these effects, we modeled Dave’s speed precisely as a function of his power, his weight, his aerodynamic drag coefficient, his bike’s friction coefficient, the road’s slope, and the wind’s strength and direction. This model, incorporating basic physical laws, is quite precise as long as we feed it accurate data.

The Internet of Dave provides some of that data. Beyond that, The Weather Company (TWC) provides wind data, and RAAM organizers provide GPS data that we use to compute slope. The following picture shows the elevation along the course:


Wind as both enemy and ally

Using our model in combination with all these data, we can predict when Dave will cross the finish line if he maintains constant power output and never rests. We ran the model with three scenarios at constant power: no wind, constant headwind of 10 mph, and constant tailwind of 10 mph.

The results were striking. In the windless scenario, Dave takes about 7.5 days to complete the race. With a headwind, he takes about one day longer. With a tailwind, he takes about one day less. As an extreme example, in 2006, riders faced the worst weather in the history of RAAM when crosswinds gusted up to 80 mph across the plains of Kansas.

Ever-changing and highly variable conditions make a clear case for using what foresight we can. So we will use TWC forecast data to compute expected headwinds and tailwinds along the route.

When to rest

Let’s suppose that Dave needs to decide whether to rest now or four hours from now. Furthermore, let’s assume that there is currently no wind but that TWC predicts a storm two hours from now that will result in strong tailwinds for four hours.

Let’s look at two possibilities:

  • Dave rests now. When he gets back on his bike, he rides four hours with a strong tailwind.
  • Dave rides now. He rides two hours with no wind, then two hours with a tailwind.


The blue curve is Dave’s speed as a function of time.

In the first scenario, Dave rides four hours with a tailwind, all at high speed. In the second, he rides two hours without wind, at a lower speed, before riding two hours with a tailwind. The first scenario clearly yields a better average speed.

Although it is easy to choose when to rest in this simple situation, picking the right time during the race will be much harder. The road isn’t straight or flat, and the wind changes continuously.

Source: Race vs. Rest Optimization (June 21st, 8 p.m. CST). With less than a thousand miles to go. The cost of resting rises substantially through the Appalachians as the net wind will be less favorable if the course were attacked after a two hour rest.

Source: Race vs. Rest Optimization (June 21st, 8 p.m. CST). With less than a thousand miles to go. The cost of resting rises substantially through the Appalachians as the net wind will be less favorable if the course were attacked after a two hour rest.

Analytics is not the decision maker

Our solution can help Dave and his crew make decisions, but ultimately it is they who will decide whether to follow our model’s recommendations. We simply assist them by providing a better way of evaluating all available options. And though compiling and calculating all that data might look like a lot of work, it’s nothing compared to the task ahead of Dave. We feel lucky and honored to be a part of his team.

Go get ’em, Dave!

If you’d like more details about the analytics and modeling we used to support Dave, please visit Analytics for the Perfect Race.

Finally, whether you’re a professional athlete in training, an amateur runner preparing for your first half marathon, or a business person who would like to make wiser decisions with your data, try Watson Analytics. It’s easy and free!