关于优化灌水的讨论

老黄

本人水平不高
灌水共大家一笑

决策基本靠感
账户基本靠攒
理论基本都乱
操作基本都反

1. max drawdown（MD) 一般不是用来优化的。 因为MD多由突发事件导致，缺少统计可靠性。MD一般用来看承受stress的能力。

老黄

2. 任何一个strategy在操作的实际过程中考虑的是当前（及预期持有过程中）的risk, 可以用trailing volatility结合市场数据等简单估算。

Diffusion

2. 任何一个strategy在操作的实际过程中考虑的是当前（及预期持有过程中）的risk, 可以用trailing volatili ...
老黄 发表于 2010-7-11 03:40

    这个有意思，展开说说。trailing volatility应该是指strategy的volatility吧，因为后文说到结合市场数据。:(13):

老黄

3. 如果系统中没有2的risk condition, 那么系统本身就不完整，优化的结果一般会seriously overfit. 2可以是简单的止损或是复杂的对冲。

老黄

这个有意思，展开说说。trailing volatility应该是指strategy的volatility吧，因为后文说到结合市 ...
Diffusion 发表于 2010-7-11 03:44

    strategy volatilities are quite often correlation with market volatility. One can use strategy vol, but better measure should be using covariance matrix among positions and risk estimation on each position. Of course, there are better ones than that too.

Diffusion

strategy volatilities are quite often correlation with market volatility. One can use stra ...
老黄 发表于 2010-7-11 03:52

不错，high correlation among positions will amplify risk。也就是说trading过程中要每天估算risk，然后根据risk来re-balance holding。

Diffusion

3. 如果系统中没有2的risk condition, 那么系统本身就不完整，优化的结果一般会seriously overfit. 2可以是 ...
老黄 发表于 2010-7-11 03:48

    能举个例子说说seriously overfit吗。我觉得overfit的一个主要来源是交易次数少，而且每个交易平均周期长。因此一两个成功的交易可以significantly change the whole output。那么优化过程很有可能会curve fitting到这一个或两个交易。有时候就算只有一个参数，也没法逃过overfit。

所以书上介绍用exit on the 5 day's close来评测一个entry strategy，应该也是为了reduce overfit...

老黄

4. carefully thought out strategies sometimes only use point in time optimization (maximize return conditon on risk) then take sharpe and alike to evaluate stragegy effectiveness. The parameter optimization is quite often run through a rolling window. The target can be total return since the first optimization will hopefully keep risk in control. This first optimization may not be necessary as long as there is a cap on estimated risk.

老黄

能举个例子说说seriously overfit吗。我觉得overfit的一个主要来源是交易次数少，而且每个交易平 ...
Diffusion 发表于 2010-7-11 04:11

    there is no way around this unless you have unlimited amount of data. this is why curve fitting method is not popular in wall st. For some high frequency trading, one probably can assume the data size id big enough.

老黄

能举个例子说说seriously overfit吗。我觉得overfit的一个主要来源是交易次数少，而且每个交易平 ...
Diffusion 发表于 2010-7-11 04:11

    theoretically speaking, using a simple (low VC dimension) rule/model  can reduce overfitting
however for marketing time strategy, simple models do not work well.

Diffusion

theoretically speaking, using a simple (low VC dimension) rule/model  can reduce overfitti ...
老黄 发表于 2010-7-11 04:20

    老大够专业啊，都是我不懂得词。看老大的帖还要另开个wikipedia才行。

Diffusion

4. carefully thought out strategies sometimes only use point in time optimization (maximize return conditon on risk) then take sharpe and alike to evaluate stragegy effectiveness. The parameter optimization is quite often run through a rolling window. The target can be total return since the first optimization will hopefully keep risk in control. This first optimization may not be necessary as long as there is a cap on estimated risk.
老黄 发表于 2010-7-11 04:11

    这里没看懂。Is "point in time optimization" a category of optimization algorithms? 没查到相关条目。

What is the first optimization? Assume the subsequent optimizations are applied on a moving window over time。

关于moving window倒是有所耳闻，也有人叫non-parameter strategy，因为不需要，parameter都是在最近的moving window上面optimize出来的。

Diffusion

theoretically speaking, using a simple (low VC dimension) rule/model  can reduce overfitti ...
老黄 发表于 2010-7-11 04:20

    看了一下VC dimension的定义。Low VC dimension在理论上成立。可是用到trading strategy optimization就很难说得通了。因为input data include market price data, which is already of very high VC dimension I guess...

Diffusion

strategy volatilities are quite often correlation with market volatility. One can use strategy vol, but better measure should be using covariance matrix among positions and risk estimation on each position. Of course, there are better ones than that too.
老黄 发表于 2010-7-11 03:52

    另外能具体说说这个"risk estimation on each position"吗？Value at Risk?

老黄

这里没看懂。Is "point in time optimization" a category of optimization algorithms? 没查到相 ...
Diffusion 发表于 2010-7-11 04:30

    'point in time optimization' means 'optimize position sizes to have high expected return under risk limit'

if you trade only SPY, then it simplifies to max SPY size under risk limit.


if you don't have this step, it will be hard to understand/control true strategy risk, because a model could take excessive risk at some point and be lucky. Another way of thinking is that one impose risk control at every point in time, thus reduce degree of freedom which will reduce overfitting in addition to better control of risk. Hope this will also answer your previous question

老黄

另外能具体说说这个"risk estimation on each position"吗？Value at Risk?
Diffusion 发表于 2010-7-11 04:39

    you can think it as volatility of each position.

老黄

看了一下VC dimension的定义。Low VC dimension在理论上成立。可是用到trading strategy optimiz ...
Diffusion 发表于 2010-7-11 04:33

    just take it as degree of freedom from physics point of view.

老黄

下网了，有问题去我屋里Q&A留言

Diffusion

没分加了...

氢氧化钠

赞老黄！胡同这地藏龙卧虎的太多了