Darryl’s Journal

[Nick Radge]

Quote:

the best results i could get so far in terms of CAR/MDD and CAR are produced with 2 lower low days over the Russell 3000. It would be interesting to know if anyone has experience with trading such a high number of signals how much effort is required!

You need to tread carefully here if you’re not using any ranking. The larger the universe and lower the entry gate, the higher the variance of outcomes

[Anonymous]

Nick Radge wrote:

You need to tread carefully here if you’re not using any ranking. The larger the universe and lower the entry gate, the higher the variance of outcomes

Ok… I’m using a ranking of: positionScore = RSI(100) but really don’t understand why this seems to give the best results. I think i understand that relative strength relates to stock being over bought (RSI ~ 70) and oversold (RSI ~30). So if I’m using a long only MR system then intuitively I would expect lower RSI values to give better results as the market picks up again. Really not sure what is going on here…

[Nick Radge]

It’s then most likely due to increased volatility of the smaller caps.

[TrentRothall]

Hi

Have you tred using Rate of Change? or maybe try RSI with a shorter look back period, then it might work better with a lower value. I think something like

PositionScore = -1000 + ref(yourRSI,-1);

should work i think

[ScottMcNab]

Be interesting to test:

PositionScore = Ref(ATR(10),-1)/Ref(C,-1);

[Anonymous]

SMcNab wrote:

Be interesting to test:

PositionScore = Ref(ATR(10),-1)/Ref(C,-1);

Results even better!

[SaidBitar]

I do not understand why you need to set position score on mean reversion system, for example if you run the exploration and you get 40 signals you will place 40 orders and will check which ones will be filled. maybe the ones will lower position score will be filled earlier than the ones with higher position score no one knows.
I think it is better to set position score as random.

[Anonymous]

[email protected] wrote:

I do not understand why you need to set position score on mean reversion system, for example if you run the exploration and you get 40 signals you will place 40 orders and will check which ones will be filled. maybe the ones will lower position score will be filled earlier than the ones with higher position score no one knows.
I think it is better to set position score as random.

what if we get 200 signals… Do we place 200 orders?

[SaidBitar]

Yes

[SaidBitar]

Honestly placing the orders is the easy part you can use the DDE that Nick talked about and all of your orders are placed

What is the hard part is after you have 200 orders how to manage them
I mean how to close all the orders when you reach your targeted number

This is why smaller universe will not be a bad idea

[Anonymous]

[email protected] wrote:

Yes

Yikes! just got the IB DDE excel spreadsheet working with the paper trading account which will help to test it for a while…

[Nick Radge]

Using a ranking system needs careful consideration and is not appropriate to all systems.

WHY
If you have too many signals and not enough cash then you will have a variance of possible outcomes, known as Selection Bias. To view the impact of Selection Bias we use the trade skipping MCS code.

The more possible signals generated means the higher the seed you should use. 10,000 signals in 20 years is starting to border on a lot. I have seen one system so far that generates 29,000 signals in that period.

If the variance of returns using a 25% seed are reasonably tight then ranking is not necessary.

However, if the variance of returns is wide, then you need to tighten that up somehow by lowering the number of signals. This can be done using price limiters, volume limiters or selecting a small universe or all of the above.

Also a ranking system is also appropriate…

…HOWEVER

Ranking will only work when you have a certain entry the next day, i.e. a market order or a definitive entry

When using a stretch, which most seem to be, then an entry the next day is NOT a certainty and therefore you may dilute your returns because you’ll only place n-orders of which y will get filled.

The absolute MUST is that your explorations should reflect the positions in the basktest. If you do an exploration and the ranked trades do not reflect the backtest positions then your real time results will not be accurate.

Personally I am not a fan of ranking and don’t use it in my own mean reversion systems.

[ScottMcNab]

How do you arrange the backtest so that the system takes 20 trades one week and 200 trades the next week ? Do you use the historical maximum number of trades ? Does this reduce exposure ?
Many thanks
Scott

[Stephen James]

I very much doubt that is possible in AmiBroker from a coding / testing perspective.

[LeeDanello]

Wouldn’t call this a stretch but this is a similar concept.

CP = Param(“Close %”,40,1,100,1);
Spread = H-L;
WeakClose = C <= L +(Spread * CP/100);

[Author]

Posts