Its that time again fuckers, lets talk algo trading, machine learning in finance (if you want to) and general quant stuff (sans greeks because that shits just boring). Stuck? Need help or advice with your algo setup? I've got fuckall to do but watch my models and wait for another round of optimization to run so ask away.
Algo trading
Grayson Harris
Other urls found in this thread:
arxiv.org
youtube.com
link.springer.com
twitter.com
Joshua Williams
what software is that, your own?
state space models are the only approach that makes sense with time series IMO, and I assume that's what I'm looking at here
>Non-Gaussian Deep Unbiased Noise SSM + PSO
Non-Gaussian so you use sampling? Is that the PSO part (particle swarm optimization I assume), or do you use something like a regular particle filter + PSO optimization?
Hunter Gutierrez
Interested. How do I start?
Jayden Nelson
> what software is that
its just a modular panel I coded ages ago to inspect models/check live performance, etc
> state space models are the only approach that makes sense with time series IMO
more true than your probably realize
> Non-Gaussian so you use sampling
> or do you use something like a regular particle filter
This is a kalman like, robust unbiased FIR filter with a non-normal noise/error term. Its basically a rewrite of the kalman filter but specifically targeting financial series. PSO is used for automatically fitting/re-fitting the covariance and the system noise matrices periodically. Dynamic decomposition (among other things) is used to check model stability.
Particle filters can and do work as well but are very sensitive to outliers/wrong model assumptions.
Alexander Watson
However, gaussian error emission is not completely useless. You could for example use a standard kalman filter, represent your model as random walk + drift, estimate noise (either through optimization or observed), draw confidence bands around the estimated mean and then act on that. You could use this to detect structural brakes (price process ceases to be AR1) or as a type of bollinger bands/mean reversion indicator. In other words, you would use the filter's error term as a trade signal
Michael Brown
Whats your maths background?
Nolan Wright
>robust unbiased FIR filter
cool I didn't know about that one, I'll implement and add it to my ssm library
what's the benefit over something like the unscented kalman filter with a cleverly chosen observation model?
>PSO is used for automatically fitting/re-fitting the covariance and the system noise matrices periodically
how do you choose the period? have you looked into online estimation of the system noise?
Logan Carter
Up to Calculus. Pretty shit. But would love to learn algo in finance.
Joseph Edwards
man, were you my colleague? we implemented exactly that where I used to work
Gabriel Richardson
well you cant assume gaussian error emission/noise in finance. So while the unscented filter is nonlinear for example, it still operates under above assumption. An unbiased iterative estimator could dismiss noise by default (which comes with its own set of problems of course). What I have above switches from normal to non-normal error emission (student's-t, etc).
> how do you choose the period? have you looked into online estimation of the system noise?
Online estimation is usually not needed once the filter converges to stable state. These types of problems arise because again the noise falls outside of assumed distributions. Online estimation is def possible though.