Thursday, March 2, 2017

Aspiring Hari Seldon - Part 2

Been quite a while, but I have a follow up to this old post.

A Crash Course

Tinkering with prices is difficult, and most players may not understand why.  Though we all interact with the price of things, unspinning the mess of how and why becomes complicated fast.  

What's worse, the price of a thing doesn't follow traditional fitting tools, it's a collection of ups-and-downs day-to-day.  It's a solution to a complicated network of factors.  This is why I've had so much trouble designing forecasts; because starting point is critical, and the walking methods aren't strictly obvious.

After dancing around this problem for as long as I have, I've come up with a few best-practices for approaching Prosper's economic reporting:
  1. Make it normal or linear: hard math is hard, keep analysis as simple as possible
  2. Figure out connections: storytelling can quickly connect seemingly disparate points
  3. Assume players aren't dumb: everyone is trying to win
  4. Look for disproof: try to eliminate weaknesses and errors in tools

Prophet - A New Crack At Forecasting

My boss sent me a link to Prophet.  And, of course, I threw EVE data at it!
On the one hand, that PLEX prediction is pretty hilarious, while on the other, the injector forecast is very close to one I'd publish on the show.

So, what is going on here?  I still think we're running into the issue of putting linear-style modeling on a non-linear problem, but we're getting much closer to the gut-version I would like to illustrate on the show.  

What's more, we can use a previous lesson and use the forecaster to predict the day-to-day volatility as a second opinion.  Using a GBM-style method, we get this:
Getting a second opinion in this case is a good way to counteract the problem of runaway models; forecasts that get stuck in a runaway up or down swing.  Though it's not fool-proof, we're reaching a "good enough" level to actually start considering it in our tooling.

Let's Go!


As promising as these initial findings are, I still worry they aren't a good replacement for more robust methods.  I'm still investigating the following:
  1. Single-dimension: only using price data, not including volume for supply/demand accounting
  2. Limited test scope: only run on CREST history data, not extended history data yet
  3. Time series rigidity: designed for daily data
  4. No extended grading done: need to test predictions vs reality
But the future is promising.  There are some key features I'm loving in this library:
  • Built-in changepoint trending: predicting discontinuities is very powerful
  • Python/R library parity: easy development/testing
  • Built-in week/month/season accounting
  • EXTREMELY FAST
There's so much more to play with, and I'm excited to dive deeper.  I'm looking into including these forecasts sparingly into the Prosper show going forward, as a better way to illustrate my gut feelings.  And hopefully we can incorporate these forecasts in a more robust feature in the near future!
TypeNames redacted while debugging


Subscribe on Twitch and YouTube

No comments:

Post a Comment