Brent’s Journal

[Anonymous]

December was my first full month of systematic trading. I am running MOC systems in both the US (R1K universe) and AU (ASX 300 universe). Pretty boring month overall.

US MOC: -70bps.

I slipped ~60bps because of two factors –

1) I entered the wrong trade file into my batch trader one day and took positions that should not have been taken. I’ve added some steps to my daily routine to avoid this in the future.

2) Before I realized that you need to sync premium data everyday, my system generated an errant price and I took on a position I shouldn’t have taken. Lesson learned…sync PD daily.

Despite the slippage, my real time results for the month fell in the middle of my month-long MCS simulation results, so I am happy with that.

AU MOC: -1.7%

Slippage here was actually a big issue. I slipped 2%, which means I would have been positive on the month if not for that. The cause of the slippage was on 12/9 I bought SRX.asx on the huge gap down, but because the opening gap was so large, I ended up buying about half the position that I should have bought. The stock rebounded well and was a big winner, but my sizing was too small relative to the intention of the system. Not much I can do about this if I don’t want to be glued to the screen.

Overall I feel OK about the month. That amount of slippage in the first month of trading is a bit disheartening…indeed if I slip this much every month my real-time results will end up being pretty underwhelming. The two issues in the US are controllable, and I am hoping that the AU issue is more of a fluke that will get washed out over time.

Putting the finishing touches on a longer-term momentum strategy that I hope to deploy in both the US and AU later this month.

[Anonymous]

January results –

I am still only trading my MOC in the US and AU at this point.

US MOC: +1.1% in Jan, cumulative (Dec/Jan) +40bps. Since inception, tracking in the 36th percentile of the MCS distribution. I had some implementation issues in December that are skewing this down…for December the result was in the 62nd percentile of the MCS distribution.

AU MOC: +91bps in Jan, cumulative (Dec/Jan) -70bps. Since inception, live performance is in the 73rd percentile of the MCS distribution.

So far so good.

[Anonymous]

Still only trading MOC in US and AU for the month of February, though I launched a US rotational strategy on March 1st. Hoping to have a longer term trend/momentum strategy working in AU by the end of the month and then I’ll have four systems live – one each short term/long term in AU / US.

Feb results –

US MOC +1%. Slippage was positive for the month at 10bps (slippage worked in my favor for 10bps). For the month, my live result fell in the 77th percentile of my MCS distribution. Since inception (late November ’16), I am +1.4%. Since inception, live results falling in the 27th percentile of the MCS distribution. Still in the lower portion of the MCS since inception due to a lot of slippage when I first started trading.

AU MOC -50bps in February and -80bps since inception in December. 52nd percentile for the month and 71st percentile since inception relative to the MCS distribution.

[Anonymous]

March results:

AU MOC: traded a partial month (shut it off after 3/22) +1.6%

US MOC: traded a partial month (shut it off after 3/21) -2.6%

Everything was going nicely until 3/21 – first true test of the systems with some volatility. Turns out I still have too much selection bias. I trade up to 40 positions and there were 80 positions possible on the 21st in the US. I could not get a result as low as I did on the 21st in a 500 run MCS. I am trading the same strategy (more or less) for the MOC in both the US and AU so I shut both off until I can do more research and fix it. Given what I am finding, I am guessing I am going to need to go back to the drawing board and completely make a new system as so far I am not finding a way to adjust the current one to bring selection bias down enough without bringing results down to ~10% CAR, which isn’t enough to trade the system.

I thought I had the selection bias figured out, since only 4.5% of days saw selection bias at all in testing, but it turns out even with such a low frequency of selection bias days, my signal capture (69%) is still too low. Nick’s post to the LMS on selection bias is timely – will need to get something with under 95% of days selection bias and over 90% signal capture I am guessing.

My live results in the US on 3/21 were 33% worse than the worst case scenario in a 500 run MCS – simply too much to trade.

I have been using a reverse channel with an ATR stretch. Any new starting points and I am all ears! I will probably start over with the R&D process on that this coming weekend.

US momentum (monthly rotational) was -2% for the month – bang in line with the simulation – very easy strategy to trade!

[ScottMcNab]

I think starting with Nicks systems of 3 lower lows (is it called second income or is that another one?) is a great basis to start from…then need to then try and modify it so not trading same system as the rest of the world ….maybe mix it in with RSI or atr filters etc

[JulianCohen]

Have a look at ROC as well. I built a reasonable MOC system using ROC < 0 to find a share that is retracing....

[ScottMcNab]

Other thing I just remembered that helped me was all the posts that Said had in his progress journal for a MRV…. with different entries, exits…so if you get a spare hour its probably worth going through that with a pen and paper and jotting down some ideas to test out as inspiration comes along

[Anonymous]

Thanks guys…time to go into hibernation / hyper-research mode! I’ll let you know how it goes!

[Anonymous]

Over the past several weeks, I have revisited the R&D stage for finding an MOC system in the US that yields meaningful results without significant selection bias. The impetus for this search was my last MOC failing when volatility reared its head in mid-March and seeing for the first time what selection bias did to my live results and how significantly they diverged from my MCS results – not pretty.

I have defined a successful MOC system as one that gets 15-20% CAGR with a MDD lower than 15% since 2010, with fewer than 5% of days having the possibility for selection bias, and at least 90% signal capture; I am looking for these criteria to be met with a 20% trade skip and I have been testing on variations of the Russell 1k, S&P 500, and S&P 100 universes.

Unfortunately, I am still not getting results that meet my criteria. I am more than a bit befuddled, having tested many different strategies and not getting anything to “work,” and thus I wonder if my approach is flawed. I question what I am missing and how I can approach the problem in a different way.

I easily design systems that show very strong results, but when I look at the selection bias, I am finding signal capture rates in the 60% range – unacceptable. When I take steps to reduce the selection bias, the results come way in. The best I am able to do with absolutely no selection bias is CAGR of ~8% on a 200% LTV system (unacceptable); with improved, yet still moderate selection bias (75% signal capture, no selection bias on 95% of days), I am able to get results around 18% CAGR / -10% MDD, but this is with no trade skip – add in slippage and results of selection bias and I am afraid the real time results will be closer to the 8-10% range, again unacceptable.
Here is the R&D process I have used thus far and some basic frameworks that I have tested – for (1) the group’s edification (e.g. this doesn’t work) and (2) if anyone can see/point out what I am missing in my approach then that is great.

Everything I have worked on has been price, indicator or price/indicator based. While I have a solid understand of indicators, I think where I am lacking is in my creativity of application – the traditional applications do not seem to work.

Here are some ideas/variations I have tried:

Basic frameworks:
• Reverse channel
• Dual reverse channel
o E.g. below 5 day low, but above the 10 day low, with variations in lookbacks
• Reverse channel w Bollinger band
o Below a channel but above a Bollinger band
o Below a channel and below a Bollinger band
• Reverse channel and ROC threshold
o Below a channel, ROC above a threshold
o Below a channel, ROC below a threshold
• Reverse channel and RSI
o Reverse channel and RSI below a threshold
o Reverse channel and RSI above a threshold
• Reverse channel with a signal count threshold – e.g if too many signals, don’t trade at all
• Reverse channel plus range expansion
o Below a channel and price range expanding over lookback
o Below a channel and price range contracting over a lookback
• Reverse channel plus historical volatility as measured by ATR:
o Above threshold
o Below threshold
• ROC below a threshold
• ROC below a threshold and not below a channel
• ROC below a threshold and an intraday reversal
• ROC below a threshold but stock outperforming index
• ROC below a threshold and stock underperforming index
• ROC below a threshold and historical intraday volatility (ATR)
o Above threshold
o Below threshold
• ROC below a threshold but not a 5 day low
• ROC below a threshold and RSI oversold
• ROC below a threshold and RSI NOT oversold
• ROC below a threshold and a volume climax
• ROC below a threshold and NOT a volume climax
• Consecutive lower lows
• Dual ROC – ST ROC below a threshold, LT ROC above a threshold
• Variations of ATR expansion
• Similar variations as reverse channel and ATR using Bollinger bands
• Using summed scores
o E.g. Sum of RSI over past three days above/below a TH
• Looking at percent rank of ROCs and RSIs, etc.
o E.g. – if RSI(5) is in the lowest 10% reading relative to itself over the past year, buy set up
• All sorts of candle patterns
• Other Factors:
o varying lookbacks
o varying ATR stretch
o varying universes
o adjusting the universe (using price constraints) based on signal count
o adjusting the universe (using universe definition) based on signal count
o eliminating trades when there are sequential set up days (e.g. eliminate signals when there has been a signal for 3 days in a row
o incorporating “puke bars” (days when prices move aggressively lower and close on lows on volume)
o ADX filters
o Volatility filters

The one thing that I have found to work well at reducing selection bias and improving system efficiency (win rate, profit factor) is layering a “medium term MA slope” filter on top of whatever other rules I am using. For me, this looks like the following:

MedTermMA = MA(C,Param(“MidTermMAPeriod”,25,5,100,1));
MAROC = ROC(MedTermMA,25);
Cond3 = MAROC >= 0;
So here we are first taking a longer term moving average (longer term relative to the primary buy criteria lookback, like a 5 day channel or 5 day ROC), then taking the rate of change of that longer term moving average, and then imposing a rule where that rate of change needs to be greater than 0. This seems to help with some of the signal clustering without doing too much to kill system results.

Using an ROC based system with my MedTermMA rule as well, here are the results I am getting and this is the best I am able to do thus far.

In the US I am testing on R1k, px filter 10-100 with 40 positions at 5% each (200% LTV).

CAGR 17.3%, MDD -6.5%, MAR 2.7, trades 13337, signals 18000, 6% selection bias days and 74% signal capture

In AU, I am testing on the ASX300 px filter > 1 and turnover >1m, 15 pos @ 13.3%.

CAGR 20%, MDD -10%, MAR 2, 4332 trades, 4751 signals, selection bias days 2.5%, signal capture 91%

The quest goes on!

[ScottMcNab]

Aus system looks nice…I have used oc>1 and turnover 500k but depends size of your account if this acceptable

US …how does it go on SPX with 20 at 10% ?

If use same equity to trade both US and AUS then can you not aim for a CAGR of 10 – 12% for each system to then combine them to meet your goals ?

For me the secret sauce was in varying the ATR stretch depending on different market conditions but not having a market filter…just make stretch bigger in certain conditions…systems does better in falling market but works in both

mine works lot better if select times of higher volatility..i just used atr

[ScottMcNab]

Out of interest Brent, how much worse was 3/21 compared to the “average” return for that day ? If it is 1% and only happens 3 or 4 times a year then may be ok ?

[SaidBitar]

wow Brent you tested everything
here is one idea that will help in reducing the signal and keep the results decent.
just pick any of the above tested idea the simpler the better and add one more filter to your system. this filter will check the sector index if it is showing pullback (such as lower BB) then you can have entries only on the stocks that are in this sector.

so in this case you are using the sector itself as filter and this will lower the selection bias in a huge way
try it and if you need any help plz tell me

[ScottMcNab]

Is this on the right track Said ?

GICS10Index = Foreign(“$SPXE”,”C”,True);
GICS15Index = Foreign(“$SPXM”,”C”,True);
GICS20Index = Foreign(“$SPXI”,”C”,True);
GICS25Index = Foreign(“$SPXD”,”C”,True);
GICS30Index = Foreign(“$SPXS”,”C”,True);
GICS35Index = Foreign(“$SPXA”,”C”,True);
GICS40Index = Foreign(“$SPXF”,”C”,True);
GICS45Index = Foreign(“$SPXT”,”C”,True);
GICS50Index = Foreign(“$SPXL”,”C”,True);
GICS55Index = Foreign(“$SPXU”,”C”,True);

pass1 =
(InGics(“10”) AND GICS10Index (InGics(“15”) AND GICS15Index (InGics(“20”) AND GICS20Index (InGics(“25”) AND GICS25Index (InGics(“30”) AND GICS30Index (InGics(“35”) AND GICS35Index (InGics(“40”)AND GICS40Index (InGics(“45”) AND GICS45Index (InGics(“50”) AND GICS50Index (InGics(“55”)AND GICS55Index ;

Cheers

[JulianCohen]

Every time I think I am happy with my system Said comes up with something that I just “HAVE” to test.

Another weekend in front of the computer….sigh

[ScottMcNab]

Apologies for the typo..NOT all meant to be Ref(GICS10Index,-1) …need to update GICS of course…

Me too Julian…so far have spent a “business” breakfast and a violin lesson constantly googling on my phone since I read this at about 5.30 am…let me know if you find a more elegant way or if this appears incorrect

[Author]

Posts