The Race To The Bottom Begins
My friends, these are the times that try men’s souls. I believe we are starting “the race” to the bottom (or not).
“We’re seeing panic all over the markets right now,” said Javier Barrio, head of equity sales for Spanish clients at Banco BPI SA in Madrid. “Governments are taking steps to try to reduce investors’ fears but confidence is weak.”[via Bloomberg]
The scary thing is that one of my simulations show the S&P500 could reach 700 rather quickly.
October 6th, 2008 at 6:48 am
Hello Tom,
A couple of questions :
1) Do you have an estimate of the accuracy of your S&P 500 model on a Test set?
2) How do your models compensate for random events?
Many Thanks!
Harry
October 6th, 2008 at 7:08 am
Harry:Yes to Q1. I use several models that either predict outcomes or that act as indicator (not prediction). The S&P500 timing model is based on bull and bear market patterns (trained by ANN’s) which include “outliers” as it tries to gauge market fear and greed. It’s similar to the VIX but better IMHO.
Q2: : Market outliers can skew my predictive models quite a bit so I use them in combination with my timing models. If the timing model shows that the market is behaving then my predictive models can have very high accuracy. If the market is turbulent, I tend to rely on my timing model more.
October 6th, 2008 at 1:29 pm
Hello Again Tom and thanks for your reply…
Actually i was wondering for a specific accuracy percentage (say 65%), so can you disclose the *actual* accuracy you get on unseen data??.
When you say about “indicators and not prediction” you are talking about numerical prediction (=regression) versus class prediction (=classification) am i right?
Last question : On your tutorial fol gold, you use several indices to predict the gold trend as LOW,RANGE,UP etc Do you use a sliding window for the class label or not?
Sorry for the many questions and Thanks Again!
October 6th, 2008 at 2:51 pm
Harry: The accuracy varies from model to model. For example, I can disclose that my volatility prediction models are about 65% accurate due to the “stickiness” of volatility. Other models, I try to get higher than that. The accuracy will vary depending on how many outliers there are. Rapidminer has a great data pre-processor that allow you to “clip out” outliers automatically as your data gets read into the model.
I used Rapidminer ANN operators to analyze bull and bear market patterns and assign a numerical indicator to represent the severity of a sell off or rally. As I stream in my market data I can see how my indicator rises or falls in real time. That indicator is then analyzed against a separate evolving genetic algorithm and standard deviations to gauge how severe the event is to tell me when to buy or sell.
Last answer: shhh, I can’t give away all my secrets.