Applied Nate Silver – another prediction failure

It’s that time again – time to check previous prediction and make a new one. The short of it is that the data-driven prediction (both short- and long-term) was wrong and my somewhat reduced expectations “gut” prediction was also wrong. More and more I’m beginning to see there is so little correlation with any measurable factors that perhaps prediction isn’t really possible, at least anything detailed. So let’s start with the money graph:

weekly18-money

So this is the second worst week in my 18 week history of this data series. I expected the week to be worst than usual but the result is even worse than expected (or data-driven predicted). Was it lack of exercise? – No, since this week is the highest recorded in last six weeks. Was it huge pigout? Yes, sorta, but really only for one day and the gap is way bigger than can be explained by the Valentine’s Day celebration.

The measured weight loss of Δ1.2lbs accounts of calorie deficit of 4200 calories for the week as a whole. But my recorded exercise was 9009 calories which implies without exercise I would have had a weight gain. That implies my daily calorie consumption was above my sedentary requirements (2176 calories/day according to various calculators). So even assuming that some of the recorded exercise was “junk” calories (very low intensity so machine records calories but didn’t really burn them, so reduce to 90% of measured) this implies daily intake was 2734. OTOH, I recorded an average of 1961, which then implies that I under-recorded by 773/day. Now that is a lot and seems seriously unlikely. Furthermore on my one pigout day I recorded 2992 and it seems unlikely it could have been much worse than that on that day (even 773 more for that day is hard to believe) so applying the entire discrepancy to consumption just doesn’t seem to fit. But if I believe the BodyMedia burned measurements (vs sedentary requirements) I only account for about 400/day. Understating consumption by 373 is more more realistic, but the whole point of all this adjusting is that I have to do a lot of “fudging” to get the data reconciled and that much fudging seems unlikely. Let’s try to check the error bounds on this:

weekly18-spreads

Adding this week’s data adjusted the trendline and almost pushed last week off the trendline and the trendline is just barely in the bounds of this week, so I still think I have about 3000 calories for the week that I can’t explain, not in scale error, not in exercise error, not in consumption error, so it makes me wonder if some new phenomenon is at work here, which means this upcoming week will be fairly critical to nail this down. (A possible hypothesis that I can observe is that I have relatively small amount of obvious body fat now but a lot of excess skin, so potentially the shrinkage of skin is much slower regardless of calorie deficit – if so, this will be an increasing trend to see in future weeks).

weekly18-deficitVSactual

It’s not much data but now I have five data points to compare the calorie deficit (as I’m generating via BodyFit measurements and recording calorie consumption) and the actual results (which have to be the “gold standard” despite some error in their measurement as well). What this shows is a relatively poor correlation (if you can even call it that) so that these measurements are so weakly correlated as to be almost useless in prediction.

So where does this leave me?

need some help here, Nate!

(I guess Nate is really busy because he hasn’t gotten back to me yet on my previous questions, but I know he’s out there so I’ll keep appealing)

The BodyMedia approach is a poor predictor and the longer-term statistical approach is a pure predictor and my “gut” is a poor predictor (I’ve been both too optimistic and too pessimistic), so what do I do – just treat this as random and unpredictable?

The problem with concluding I can’t predict will come when I try to move into maintenance. Since I’m still in aggressive loss mode (although it’s getting harder and my commitment is weakening) a missed prediction can be made up in the future, but once I’m at my minimum (now looking further in the distance) observing any uptick and quickly compensating is going to be challenging. The 9009 recorded calories of exercise is way more than I can sustain long-term (even 4000 is fairly optimistic) so that implies my food intake is about 700 calories/day too high and that’s awful because subjectively my intake was very modest (way way below what it used to be) nor was it excessive in the “quality”. In fact, if I were following what the nutritionists and dieticians say I would have had a significant gain this week. So my partial starvation this week will have to be severe starvation for the rest of my life, esp., after I can no longer sustain this level of exercise. That’s rather gloomy.

Soon (after 8 weeks from now) I’ll be scheduling events that will trigger significant deviation from plan so it’s entirely possible April will be a net gain with most of May then spent recovering from that and then June has another be uptick in it so I might still be looking at my birthday before I achieve my goal. I’m not sure I can handle that much negativity and still maintain my resolve, so let’s hope this week was an anomaly somehow.

So with no expected deviations from plan this week I’ll predict a conservative value:

205.0

which will finally put me at a total Δ of 50lbs, which isn’t bad, but still not enough (20lbs more to go after that). So next week, I may attempt a longer-range prediction.

p.s. While proofreading this post I did another data analysis. If I look at the five full weeks of BodyMedia data I achieved a daily deficit of 1385 calories but if I look at weight loss over that time (converted to calories) the deficit is only 1070. So BodyMedia’s app overstates by  about 30% so in general I should reduce the amount that shows in the spreadsheet to about 75% to get a better guess (my “gut” adjustment was more aggressive, 60%, but the weak correlation shows the problem there, so it will take several more data points to begin to get an evenly vaguely accurate adjustment).

Advertisements

About dmill96

old fat (but now getting trim and fit) guy, who used to create software in Silicon Valley (almost before it was called that), who used to go backpacking and bicycling and cross-country skiing and now geodashes, drives AWD in Wyoming, takes pictures, and writes long blog posts and does xizquvjyk.
This entry was posted in comment, musing, prediction and tagged , , . Bookmark the permalink.

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s