This thread is intended to be a serious version of http://elitetrader.com/vb/showthread.php?t=227873 without the flame war.
So I think we all understand (due to atticus' quote, and sle, and others...) that you need to have some kind of edge in predicting vol and/or direction to be profitable. My question is, how do you guys do it? I am not asking for the keys to the castle, but what sort of modeling approach do you take?
- Is Sinclair right that every options trader basically thinks of the world as a GARCH model?
- Do you extend this to integrated/fractionally integrated processes?
- Do you only model vol? It seems a common assumption that drift is zero (e.g., through the use of log returns), and therefore every move is vol, but I find that hard to believe in some cases.
- How do you model correlations/dependence between names?
- How do you cap the number of lags/meaning of lags when the data are higher frequency?
- Are there a few key insights or key academic papers that lay out a useful model selection process? (i.e., a process that avoids over-fitting)
- Perhaps most important, what sort of predictions are we looking for the model to make? That vol relationships between names will mean revert? That skew/term structure in one name will converge to a long term relationship?
On a related note, do you find that markets are sometimes predictably mean-reverting? Based on a model of markets as a continuous-time bidding/auction process, it seems to me they would have to be, save for a percentage of market participants acting on trend-following strategies. If so, it would seem that no-arbitrage arguments require all options to be overpriced from a directional betting perspective, because mean-reversion would make their delta-hedged value higher than their value based on the distribution of prices at expiry. If we are comparing delta-hedged replication vs. distribution of terminal prices, then drift now plays a role, too. Thoughts?
mj
So I think we all understand (due to atticus' quote, and sle, and others...) that you need to have some kind of edge in predicting vol and/or direction to be profitable. My question is, how do you guys do it? I am not asking for the keys to the castle, but what sort of modeling approach do you take?
- Is Sinclair right that every options trader basically thinks of the world as a GARCH model?
- Do you extend this to integrated/fractionally integrated processes?
- Do you only model vol? It seems a common assumption that drift is zero (e.g., through the use of log returns), and therefore every move is vol, but I find that hard to believe in some cases.
- How do you model correlations/dependence between names?
- How do you cap the number of lags/meaning of lags when the data are higher frequency?
- Are there a few key insights or key academic papers that lay out a useful model selection process? (i.e., a process that avoids over-fitting)
- Perhaps most important, what sort of predictions are we looking for the model to make? That vol relationships between names will mean revert? That skew/term structure in one name will converge to a long term relationship?
On a related note, do you find that markets are sometimes predictably mean-reverting? Based on a model of markets as a continuous-time bidding/auction process, it seems to me they would have to be, save for a percentage of market participants acting on trend-following strategies. If so, it would seem that no-arbitrage arguments require all options to be overpriced from a directional betting perspective, because mean-reversion would make their delta-hedged value higher than their value based on the distribution of prices at expiry. If we are comparing delta-hedged replication vs. distribution of terminal prices, then drift now plays a role, too. Thoughts?
mj
) and never had a losing year. So, what does that prove?
