Statistics for profit

Joel Spolsky's Evidence Based Scheduling is a smart idea. Really smart.

I have a few questions about it I'd love to have answered, but there's no comments section or forums entry for this post:

• Do you make normal assumptions for the error on estimates? Seems more likely that you would expect the distribution of incorrect estimates to skew towards too long?


• Aren't the errors heteroscedatic, I'd expect the error to be larger on larger bites of data (and likewise, shorter on smaller bites)?


• Can you produce a conditional distribution for the likely finish dates rather than just a confidence interval on the expected on?


• Can you track dependencies in teh schedule, that is can we propogate an uncertainty from one milestone on one developer into the start of another milestone for a second developer, this would really help measure the interaction costs of a large team, which seems to be missing in this model?


• Why is this technology stuck inside of FogBugs rather than being its own product. There are tons of use cases for this in all sorts of knowledge labor fields, why not enable those? Also isn't in an orthogonal feature set to defect tracking?

I don't need these features, but having them would really make me happy.

I like me some precious

okay. i get it. you think Darjeeling is a bit precious.

Wes Anderson spends a lot of time on the art direction. And some of the shots are maybe a bit contrived.

But seriously, where else do you get to shear joy of the mundane parts of our lives. Name another movie that makes a fetish of luggage. (Though its also the obvious metaphor in the film)

A film about moving on in which the location moves along with you? On a pre-ordained route inscribed by steel rails and wood ties? You love that stuff. You know you do.

Wes Anderson remains films answer to Richard Powers. Forever constructed intricate and meticulously crafted structures in which no item has no place, and all the parts form the unique and necessary conditions to maintain the story, Anderson has a tendency, like Powers, to lose the soul of the story sometimes under all teh technical greatness. Just be grateful no movie is 800 pages long. Its also worth mentioning, that like Powers, Anderson might need to move on from handling the lives emotionally stunted and withdrawn men of roughly the same age as Anderson himself.

But none of that takes away from the joy of the games, puzzles and distractions his movies provide.

And did you catch that micro-movie on itunes, Hotel Chevalier? A masterpiece. Better than Darjeeling, frankly. But you can't claim all the visuals cues passed back and forth between the two didn't increase yr enjoyment of both movies.

The world has spoken and it judged Wes Anderson precious. Which I still don't get as a pejoritive so much. But I'd rather have more artists fall in love with the sound of them trying to make something than have then fall in love with the sound of their own voice.

Helvetica

How do you make an awesome movie about a font?

You start with a great ambiant rock soundtrack

Then feature a 2 minute long montage of a goofy whitehaired man running around New York pointing at signs with Helvetica in it.
Then don't so much make it about a font, but about the development from modernism to post-modernism and the subsequent rejection of post-modernism. Tell a narrative about who we are as consumers and how packaging reflects our identity.

Helvetica is all of that and more. Bringing together interviews with what I understand are some of the most important designers I've never heard of and visually telling a history of design by focussing on such a small topic that anyone (even I) can get their head around it really does wonders.

For two hours JP, SB and others enjoyed this movie and everyone came out smiling.

Check out clips here and see if its for you. Listings on on their blog.

Fun With Heat Maps

It's been a while. Long enough that RD called me out on not posting since the summer. I can't promise to keep up with his ferocious pace, but I really ought to get back in gear. So hear we go (we'll start easy):

As many of you know (and have been bored to tears by) I've been wasting a good deal of my time on the Netflix Prize. Other than failing submissions, there's a bunch of interesting byproducts of this effort. Here are a few heat maps I've created from the Netflix Prize Corpus. There's not a lot of information there, but I thought a few of them were pretty interesting to look at. Not much to say about them other than color represents probability density, and the curves on the left and the bottom of the charts are the marginal distributions of the probability density marked up for the contribution coming from different intensity sources (so a very cool colored peak would be a long tail effect, etc.) The colors suck, if someone has a good system for allocating colors to intensity someway I'd love to hear about it.

There is some pattern emergence in a few having to do with the integer nature of the data that you don't get to see every day.


Some of them need extra binning on account of the dimension variable being flakey, so there's a lot less density over all here.


A few misleading, but tempting linear relationships that you can't, actually, run a standard regression against with any hopes of getting real results.





And some that show just about perfect noise.