microprediction (Peter's Repos)

See also:

• Papers, working papers, articles.
• My thanks for reaching out page.
• Medium blog
• Linked-In content.

About me

Career quant, applied mathematician, open-source developer.

Technological interests

• I'm long the use of large language models for alpha, code gen, prediction and instrumentation (e.g. pi).
• I'm long decentralized AI and the Strong Indispensable Markets Hypothesis, which is to say I like to provoke people into using market-inspired collective mechanisms for prediction. I used the options market to effortlessly beat 97% of participants in the year-long M6 contest - see the post or article. My book is a meditation on the power of mini-markets and algorithmic micro-managers in a very specific yet ubiquitous domain: frequently repeated prediction. It predates and extends "Info Finance" coined by Vitalik Buterin. Read the awards and reviews. I run contests like mid-one which you might want to check out too as there is a lot of prize-money.

Mathematical interests

• Portfolio and ensemble construction (e.g. paper and blog where I unified the two sides of portfolio theory).
• OTC microstructure (my day job). I'm not currently focussed on credit but have been (see this or that or the other)
• Thurston models
• Derivative-free optimization
• Time-series
• Sports analytics

MontePrediction.

Allow me to mention a long-running game where you hurl a million 11-dimensional Monte Carlo samples at my server.

1. Open this colab notebook or script (yes there's an R version),
2. Change the email, at minimum,
3. Run it. Every weekend.
4. Check your scores at www.monteprediction.com

The notebook also describes the scoring mechanism. Ask questions in the slack (see bottom of leaderboard for slack invite).

Now, here's some attempt to introduce you to my open source work ...

Derivative-free optimizer comparisons

The humpDay package is intended to help you choose a derivative-free optimizer for your use case.

Schur Complementary Portfolios

My work on unifying Hierarchical Risk Parity with minimum variance portfolio optimization sits in the precise package although I've been meaning to put it somewhere else.

Incremental time-series and benchmarking

The timemachines package enumerates online methods and makes some effort to evaluate univariate methods against the corpus of time-series drawn from the microprediction platform. It is an attempt to reduce everything to relatively pure functions:

$$ f : (y_t, state; k) \mapsto ( [\hat{y}(t+1),\hat{y}(t+2),\dots,\hat{y}(t+k) ], [\sigma(t+1),\dots,\sigma(t+k)], posterior\ state)) $$

where $\sigma(t+l)$ estimates the standard error of the prediction $\hat{y}(t+l)$.

Benchmarking overview

Topic	Package	Elo ratings	Methods	Data sources
Univariate time-series	timemachines	Timeseries Elo ratings	Most popular packages (list)	microprediction streams
Global derivative-free optimization	humpday	Optimizer Elo ratings	Most popular packages (list)	A mix of classic and new objectives
Covariance, precision, correlation	precise	See notebooks	cov and portfolio lists	Stocks, electricity etc

These packages aspire to advance online autonomous prediction in a small way, but also help me notice if anyone else does.

Winning

The winning package includes my recently published fast algorithm for inferring relative ability from win probabilities, at any scale. As explained in the paper the uses extend well beyond the pricing of quinellas!

First Down

My firstdown repo contains analysis aspiring to ruin great game of football. See Wilmott paper and for heaven's sake, don't stretch out for the first down. That's obviously nuts.

Some other repos

• randomcov - A set of quirky correlation and covariance matrix generators (I'd love your ideas).
• embarrassingly - A speculative approach to robust optimization that sends impure objective functions to optimizers.
• pandemic - Ornstein-Uhlenbeck epidemic simulation (related paper)
• momentum - My most personally re-used mini package ... for incremental mean, var, skew, kurtosis.
• muid - Memorable Unique Identifiers ... try to figure out how that can't be an oxymoron.
• timeseries-notebooks - Lots of examples of using open source timeseries packages.
• correlationbounds - Mini package for conf bounds
• building_an_open_ai_network - Book related.
• recalibrate - Utils related to Platt scaling etc.

The currently defunct real-time time-series platform

The real-time system previously maintained by yours truly has entered a trisoloran dehydrated state but will hopefully be revived at a future date, after one of my three hundred ChatGPT generated scientific grant proposals is successful. Here's how some of the open-source stuff used to propagate down into the "algo fight club".

1. The "/skaters" provide canonical, single-line of code access to functionality drawn from packages like river, pydlm, tbats, pmdarima, statsmodels.tsa, neuralprophet, Facebook Prophet, Uber's orbit, Facebook's greykite and more.

2. The StreamSkater makes it easy to use any "skater".

3. Choices are sometimes advised by Elo ratings, but anyone can do what they want.

4. It's not too hard to use my HumpDay package for offline meta-param tweaking, et cetera.

5. It's not too hard to use my precise package for online ensembling.

Those repose are:

• The muid identifier package is explained in this video.
• microconventions captures things common to client and server, and may answer many of your more specific questions about prediction horizons, et cetera.
• rediz contains server side code. For the brave.
• There are other rats and mice like getjson, runthis and momentum.