Thanks. Indeed my system is not trading gas at the moment due to the high volatility. Well, it did open a short in gas recently but the cold spike kicked me out again.
Fully automated futures trading
- Thread starter globalarbtrader
- Start date
The answer is in GAT's blog
- Rob Carver22 August 2016 at 19:04
Sharpe is not statistically significant. Read chapter 3 ... again.
-
Murat24 August 2016 at 11:46
Ah. You use t_test and p_value calls on daily returns data and test if they are statistically different from 0 - in this case, p-value is > 0.05 - not significant.
So we cannot say anything for sure on CORN - but one could still use it because it provides diversification? I ask because here
http://qoppac.blogspot.co.uk/2015/07/a-tale-of-two-positions.html
you mention CORN position would have meant a worse loss during 2015 were it not for another agricultural instrument in the portfolio.
I also checked Andrew Lo's paper that shows how to compute Sharpe ratio confidence intervals,
def err(SR,T):
return np.sqrt( (1 + 0.5*SR**2)/T )
T = 40*256; sr = -0.12
print sr + np.array([-1, 1]) * err(sr,T)
gives [-0.12991763 -0.11008237], both ends for SR for CORN are in the negative: I don't know how to connect this with the previous significance results, but I guess rejection from t-test is stronger.
Thanks,
Rob Carver31 August 2016 at 12:09
The anecodote in that paper has no statistical meaning eithier - it's just one data point. But sure, Corn isn't perfectly correlated to other subsystem returns, so if you assume we can't predict risk adjusted returns then we should add it.
I'm not familiar with the test you show, so I can't comment on it.
There is also a multiple testing problem - if we test enough instruments then some of them will be significantly negative or positive regardless of the real underlying properties of the trading system.
The answer is in GAT's blog
- Rob Carver22 August 2016 at 19:04
Sharpe is not statistically significant. Read chapter 3 ... again.
Murat24 August 2016 at 11:46
Ah. You use t_test and p_value calls on daily returns data and test if they are statistically different from 0 - in this case, p-value is > 0.05 - not significant.
So we cannot say anything for sure on CORN - but one could still use it because it provides diversification? I ask because here
http://qoppac.blogspot.co.uk/2015/07/a-tale-of-two-positions.html
you mention CORN position would have meant a worse loss during 2015 were it not for another agricultural instrument in the portfolio.
I also checked Andrew Lo's paper that shows how to compute Sharpe ratio confidence intervals,
def err(SR,T):
return np.sqrt( (1 + 0.5*SR**2)/T )
T = 40*256; sr = -0.12
print sr + np.array([-1, 1]) * err(sr,T)
gives [-0.12991763 -0.11008237], both ends for SR for CORN are in the negative: I don't know how to connect this with the previous significance results, but I guess rejection from t-test is stronger.
Thanks,
Rob Carver31 August 2016 at 12:09
The anecodote in that paper has no statistical meaning eithier - it's just one data point. But sure, Corn isn't perfectly correlated to other subsystem returns, so if you assume we can't predict risk adjusted returns then we should add it.
I'm not familiar with the test you show, so I can't comment on it.
There is also a multiple testing problem - if we test enough instruments then some of them will be significantly negative or positive regardless of the real underlying properties of the trading system.
I just added this comment to that thread, as I've now re-read Lo's paper:
"Incidentally you made a mistake with Los formula - you are using an annual SR but you're assuming you have daily data. The correct range for the confidence interval given an annual SR of -0.12 is -0.43, +0.46"
GAT
Hi, I'm a long-time follower, but only now finally got around building the thend system based on the Rob's great material. Actualy I'll be extending my existing system (stocks stat arb, not greatly profitable but at least not a total looser 
), which is event-driven (I.e. I process every tick, make decisions and place/adjust orders based on the signals every time).
Hence my question (would appreciate if anyone can answer): For the carry-signals, can I (should I) evaluate the forecast in real-time or only based on the EOD-prices?
I.e. with EWMA-thend I guess it's no problem to evaluate the forecast on every tick as the "price-contract" is liquid (just add the last price to the series of the EOD prices and pretend like it's a new EOD, but don't store it in the EOD-series after evaluating the rules, the next rt-price will replace it on the next tick and only the new EOD price will be stored in the series), but with carry there could be problems: my "carry-contract" might not be liquid, therefore the tick-prices I'll be getting, which will be compared with the prices of the price-contract might not be in sync or not real. Also, we're applying smoothing to cary-signals to avoid seasonality and slow down trading (didn't fully figure out how to choose the lookback for diff. contracts yet..), so maybe evaluating carry on rt-prices makes no sense. In that case, I guess the approach could be to calculate carry only once a day and use it for combining with tick-based EWMA forecast till the end of the current day..?
Appreciate any suggestions, thanks.
), which is event-driven (I.e. I process every tick, make decisions and place/adjust orders based on the signals every time).Hence my question (would appreciate if anyone can answer): For the carry-signals, can I (should I) evaluate the forecast in real-time or only based on the EOD-prices?
I.e. with EWMA-thend I guess it's no problem to evaluate the forecast on every tick as the "price-contract" is liquid (just add the last price to the series of the EOD prices and pretend like it's a new EOD, but don't store it in the EOD-series after evaluating the rules, the next rt-price will replace it on the next tick and only the new EOD price will be stored in the series), but with carry there could be problems: my "carry-contract" might not be liquid, therefore the tick-prices I'll be getting, which will be compared with the prices of the price-contract might not be in sync or not real. Also, we're applying smoothing to cary-signals to avoid seasonality and slow down trading (didn't fully figure out how to choose the lookback for diff. contracts yet..), so maybe evaluating carry on rt-prices makes no sense. In that case, I guess the approach could be to calculate carry only once a day and use it for combining with tick-based EWMA forecast till the end of the current day..?
Appreciate any suggestions, thanks.
Last edited:
Hi, I'm a long-time follower, but only now finally got around building the thend system based on the Rob's great material. Actualy I'll be extending my existing system (stocks stat arb, not greatly profitable but at least not a total looser
Hence my question (would appreciate if anyone can answer): For the carry-signals, can I (should I) evaluate the forecast in real-time or only based on the EOD-prices?
I.e. with EWMA-thend I guess it's no problem to evaluate the forecast on every tick as the "price-contract" is liquid (just add the last price to the series of the EOD prices and pretend like it's a new EOD, but don't store it in the EOD-series after evaluating the rules, the next rt-price will replace it on the next tick and only the new EOD price will be stored in the series), but with carry there could be problems: my "carry-contract" might not be liquid, therefore the tick-prices I'll be getting, which will be compared with the prices of the price-contract might not be in sync or not real. Also, we're applying smoothing to cary-signals to avoid seasonality and slow down trading (didn't fully figure out how to choose the lookback for diff. contracts yet..), so maybe evaluating carry on rt-prices makes no sense. In that case, I guess the approach could be to calculate carry only once a day and use it for combining with tick-based EWMA forecast till the end of the current day..?
Appreciate any suggestions, thanks.
I'd certainly only calculate carry on end of day prices, since these are synchronised, and having synchronised prices is essential.
GAT
Thanks Rob!
Well, you can create a concurrent market based on the incoming events which is fairly straightforward. Of course, it mainly makes sense if you are planning to add micro-alphas to your process.
Seems sensible, good choice of instruments. For eurodollar, can I check you're trading many contracts away (I trade 900 days out)? and for v2tx, you're not in the front contract?
It's not statistically significant. I also added the rule, and saw that there wasn't the statistical evidence to include it, so removed it.
I did implement @globalarbtrader 's short bias on vol futures.