Thursday, 7 April 2016

Futures trading performance - year two

Well time flies. It's a year since I wrote this post:

http://qoppac.blogspot.co.uk/2015/04/futures-trading-performance-year-one.html

... which means it's time I reviewed my second year of systematic futures trading.

Executive summary: Last year was amazing, this year was just excellent.


The money shot




Or to put it in the context of the full period:

So basically a 40% increase, on top of the 57% last year (I don't compound my returns). Unlike last year I didn't change my risk weighting.

As I've stated before the p&l in my account comes from two main sources; a hedged stock portfolio (you can think of this as a long:short equity hedge fund), and some futures trading.

Here's the full breakdown

Stocks
Mark to market: -5.06%
Dividends 4.63%
Commissions: -0.02%

Subtotal: -0.45%

Hedge
Mark to market: 16.79%
Commissions: -0.03%

Subtotal: 16.73%

Total for stocks and hedge: 16.31%

Futures
Gross profit: 21.48%
Commissions: -0.76%
Slippage: -0.63% (Bid ask -1.2%, less execution algo profit 0.56%)
Data fees: -0.03%
FX trading: 3.22%
Interest: -0.04%

Total for futures: 23.24%

Grand total: 39.55%

The hedged equity portfolio made a lot more money than last year. It's important to bear this in mind when comparing against pure futures managers as I will do later on in the post.

My futures trading was still very good (a SR of just over 1.0; better than expected in the back test at around 0.88) though not as amazing as last years 57%.

Slippage was about half the level of the prior year; probably because my execution algo was doing a wonderful job for the whole year rather than just part of the year as before. I'll be analysing my execution costs in more detail in another post.

Total trading costs then come in at 1.39% versus a backtested 1.5%. Once again slippage was lower than expected (1.0% in backtest versus 0.63% this year) and commissions were higher (0.5% in backtest versus 0.76%).

The back test doesn't include futures rolls; I can't be precise but eyeballing the data they account for around a third of commissions and a similar chunk of bid-ask so in reality I'm probably doing okay. This will be something I'll be implementing in my new system, pysystemtrade.


Fun facts and figures


Realised standard deviation (based on weekly returns) 24.6% versus target 25%
41.4% of trades were winners
Avg profit:avg loss ratio: 1.841
Best day +7.6%
Worst day -5.4%
Worst drawdown -18%


Benchmarking


As I noted above it's probably unfair to compare my performance to a pure futures CTA like AHL, but for consistency I did it anyway. As with last year I used the AHL diversity GBP class as my benchmark. The correlation of monthly returns for the last 12 months was 0.80, a little lower than before, but still relatively high.

Because AHL runs at a lower vol than me to graph our returns I first need to adjust AHL's for this effect. This gives me the following cumulated returns for the two year period:


The Sharpe ratios for the last 12 months were 1.33 (me), -0.59 (AHL). If I take out the additional performance of equity neutral then my Sharpe is around 1.0. So just on pure futures trading I've still done much better this year. It isn't just AHL eithier. Of the big European CTA names only Aspect (+8%) and ISAM (+15%) made anything like a SR of 1.0 (Winton was up slightly).

I'd guess this is probably due to having more carry in my system (I underperformed in the first year, when pure trend following kicked butt). It could also be a lower allocation to under performing asset classes like equities (of which more in a second).

For the whole period the Sharpes are 1.65 (me) and 1.38 (AHL) respectively. Again without the equity neutral component we'd have very similar performance over the two years.

There is a longer discussion of the "big versus small" argument here.


Performance by instrument


 Last year was the year of the bond. This year, with twisting and turning in the Fed's policy causing ructions and sharp reversals in financial assets, has been the year of non financial assets.

Metals 13.4%
Oil and gas 12.5%
STIR 4.5%
Bonds 2.2%
Vol 0.1%
FX -0.1%
Ags -1.3%
Equities (ex hedge) -5.8%

The trend in all the precious metals (Gold, Platinum and Palladium) was strong, with Copper helping as well.

Let's look at Gold since most people know the chart:

Not a single one off trend; rather two down trends and one up trend on which serious profit was made; interspersed with gradual leakage. Gold actually ended up the year unchanged (with rolldown included), yet we made money. Lovely.

Crude was a winner last year, but Gas actually contributed more this year:

A more straightforward strategy here; short, short and keep on shorting...

Although bonds didn't do very much, Eurodollar's (now the only market in the STIR asset class) were a stand out:

Despite Ms Yellen's best efforts the overall direction was up; and this basically translated into profits (though I was a little flatter in the trend less periods as you'd expect).

One of the worst markets was CAC; very typical of stock markets generally:

What an ugly chart. Most of the time there wasn't a trend strong enough to justify a position; then when one was put on it was immediately whipsawed out, resulting in gradual bleeding.

However cable IMM futures were the worst market of all

Almost a year of perfectly awful rangebound behaviour; with one decent trend saving my bacon a little.


Fin


As I have noted before, and will note again, and again, and again; even a few years worth of live trading performance may not tell you with much statistical significance whether you are good or just lucky (or both). I've definitely been lucky with my equity neutral strategy this year; whilst my futures trading is in line with expectations.

At best I can say it's becoming less likely that I'm doing something stupid and have just been lucky. So if markets continue to behave as they have for hundreds of years, then I should continue to make a little money.

If you're interested in building your own systematic trading engine, then have a look at my other blog posts on the subject, or of course you can also buy my book.

24 comments:

  1. I run an equity neutral strategy in my account with similar 2015 results. +24%. However 2016 has been rough to be currently sitting -11%. Some of this i attribute to poor management on my part, but i was wondering if you would be willing to share your Q1 2016 experience?

    ReplyDelete
  2. The problem is I can't easily break out the performance of the equity hedge which is why I only do it annually.

    ReplyDelete
  3. Hello Rob,
    I have read your book, and must admit it's a very interesting reading.

    I have one question regarding your performance in the cable. Starting from around May 2015 there seems to be a steady although not steep downward trend. To what you do attribute your negative performance? Was perhaps your system too slow to reverse from a previous upward trend?

    ReplyDelete
    Replies
    1. If anything I was probably trading this market too quickly rather than too slowly.

      Filters (like EWMAC and other indicators) are good at capturing certain kinds of trends but not others. In this particular case a much simpler indicator (eg regression line) would probably have "seen" this trend and gone short. Generally however filters do a better job.

      Delete
  4. Hi Rob - tried posting this question earlier but google didnt like it.
    Anyway - big fan of your blog. Currently reading ur book.

    My question: Would you recommend systematic or semi-systematic trading when only applied to a very small universe of assets ? Or is the lack of market diversification too much of a no no for you?

    Some context:
    I'm a paper trader in a physical commodity house so i can only prop in Brent, WTI etc... I have been experimenting (in Excel) with composites of moving average crossovers with some interesting results.

    I guess you would need to develop a wider variety of systems that aren't historically correlated to diversify the strategy risk a bit since you are only applying it to a small universe of markets ?

    Let me know your thoughts (if any). Thanks.

    ReplyDelete
    Replies
    1. I've already replied here

      http://qoppac.blogspot.co.uk/2016/01/computers-vs-humans-considering-median.html?showComment=1460385977209#c8890135491984810666

      Delete
    2. Sorry - it didn't seem to register on my Google acc. Apologies for that.

      Delete
    3. Comments are moderated so won't show up immediately. I also delete duplicates.

      Delete
  5. Hi Rob, Two questions:

    In Facts and Figures above, are those stats for the overall portfolio including hedged stock or just the futures trading component?

    You have made a strong point about the need for a long span of data (35 years if I recall your suggestion correctly from the book) relating to this type of longer term trading system . How would that translate for a higher speed system that makes discrete trades daily? Is there a minimum number of samples that would be sufficient or is it really a function of proving robustness over all market conditions that might occur over a very long period of time?

    Thanks, John

    ReplyDelete
    Replies
    1. The whole portfolio.

      For faster systems you'll get more statistical evidence quicker ONLY if the sharpe ratio is higher.

      (This is because otherwise if you increase the frequency of trades or returns, then both the noise and the statistical likelihood chance with square root of time, so [ignoring autocorrelation] the t statistic will be the same)

      The big high frequency strategies have sharpe ratios north of 2.0.

      So if you look at table 5 then a couple of years data would be more than enough for these guys.

      Delete
  6. Hi Rob
    I'm sorry if this question has been covered before but perhaps not? The question is, "is there a formula to determine the forecast weights from your correlation matrix rather than estimating them from your look-up tables. I thought I could work around the problem using excel solver on a data series, but suspect there may be a neater programmable solution."
    Many thanks
    Chris

    ReplyDelete
    Replies
    1. I don't understand the question. Do you want the formula to calculate correlations, or for calculating portfolio weights given a set of correlations?

      Delete
  7. The latter - I would like to calculate the portfolio weights given a set of correlations
    Chris

    ReplyDelete
    Replies
    1. Well classic markowitz obviously, though bootstrapping or shrinkage is better

      http://qoppac.blogspot.co.uk/2015/10/a-little-demonstration-of-portfolio.html

      Delete
    2. Thanks Rob - I'll have to study that one in more detail!

      Delete
  8. Hi Rob, Are you planning to make the slides from your CMC seminar available? Thanks

    ReplyDelete
    Replies
    1. https://www.docdroid.net/8J4RVB7/mythmtalondonfinalv3.pdf.html

      Delete
  9. Interested to see April's performance update!

    ReplyDelete
  10. I see MAN AHL Diversified is really struggling lately. How are you doing?

    ReplyDelete
    Replies
    1. https://fundseeder.com/trading_account/1246 (registration required)

      https://www.elitetrader.com/et/threads/fully-automated-futures-trading.289589/page-49#post-4326241

      Delete
  11. thanks again for sharing Rob, I read your book after listening to your podcast with Niels. Answered many questions and clarified concepts i was trying to figure out the best way to implement.
    What was your margin to equity ration in 2016? also, may investors consider return after normalising for margin to equity and vol. whats your view on this?

    ReplyDelete
    Replies
    1. Margin to equity ran at an average of around 30%.

      Obviously considering margin will make strategies like mine more attractive compared to say relative value equity (lower vol, but higher margin requirement).

      Delete
  12. Rob,
    Would you say that, in general, the risk of overfitting trading rules for a strategy is (much) greater when doing infrequent trades (say 1-2x/day) with lower-frequency data (1-10 updates/day), versus >50 trades/day with tick data? Maybe the line between 'curve-fit' and 'over-fit' is gray in spots, esp when frequency of evaluation versus intraday volatility comes into play?

    ReplyDelete
    Replies
    1. You'll overfit if your model has too many parameters relative to the number of meaningful data points. So yes 50 points a day for 40 years is better than 1 point a day for 40 years. However a few caveats: (a) the data has to be meaningful, if you your turnover is the same with both models the extra data won't help, (b) it isn't going to be 50 times better - you can't put 50 times the number of parameters in, (c) usually you're replacing 40 years of daily data with much fewer years of tick data so you don't see the full benefit.

      Delete