Machine learning in trading: theory, models, practice and algo-trading - page 1142

[pantural]

Rashid Umarov:

I took your report, copied trades from it and made a calculation in Excel based on them.There is nothing complicated, look at the formulas and you will understand that Gods do not burn pots. I attach the file

As you can see, in the test report the Sharpe coefficient is counted correctly.

I repeat, your algorithm is not correct, it is a classic first-year economics student mistake to forget about the root of the sample length when calculating SR of different lengths. The value of such a calculation will be significantly different for different number of trades and it will not be possible to compare the equity for a month and year. Holy shit, google it or something... Or I will have to post these calculations on elitrader and it will be a shame, because it is not some custom software, but one of the dominant and such a fiasco...

PS SR >3 is quite a normal value of SR, unless of course it's a fit, with HFTs it can be double-digit:)

[Aleksey Vyazmikin]

Rashid Umarov:

Show it right here, otherwise you'll have to banish it.

PS It is assumed a reasonable initial deposit and lot size for trading. Not too small and not too big. And then there will be no question why an average profit of $1 on a deposit of $100 k shows a worse Sharp than the average profit of $100 on a deposit of $1000.

[Rashid Umarov]

pantural:

Once again, your algorithm is wrong, it is a classic mistake of a first-year economics student to forget about the root of the sample length when calculating SR of different lengths, the values of this calculation will be significantly different with different number of trades, it will not be possible to compare equity for a month and a year. Holy shit, google it or something... Or you will have to post these calculations on elitrader and it will be a shame, because it is not some custom software, but one of the dominant and such a fiasco...

You have a bit of a problem - first you come up with a formula and attribute it to us, then you try to find an error in that formula. Вспомните/почитайте википедию , например https://ru.wikipedia.org/wiki/%D0%9A%D0%BE%D1%8D%D1%84%D1%84%D0%B8%D1%86%D0%B8%D0%B5%D0%BD%D1%82_%D0%A8%D0%B0%D1%80%D0%BF%D0%B0

There you can also find the standard deviation at the link

[pantural]

Aleksey Vyazmikin:

If we resample (for example, double the amount of equity), their formula will be completely different, although the ratio of profit to risk has not changed much)))

[Rashid Umarov]

pantural:

This is just the beginning, with this algorithm, if this equity is resampled (for example, to prorate in half), their formula will be completely different, although the ratio of profit to risk has not changed almost))))

There's no need to use clever words. Resample belongs to another topic, here we take only PnL sample without any timeline and count Sharpe ratio on it.

[pantural]

Rashid Umarov:

You have a bit of a problem - first you come up with a formula and attribute it to us, then you try to find an error in that formula. Вспомните/почитайте википедию , например https://ru.wikipedia.org/wiki/%D0%9A%D0%BE%D1%8D%D1%84%D1%84%D0%B8%D1%86%D0%B8%D0%B5%D0%BD%D1%82_%D0%A8%D0%B0%D1%80%D0%BF%D0%B0

You can find the standard deviation at the same link.

Look, Wikipedia is no substitute for WO, Wikipedia doesn't take into account the whole gamut of what happens in practice, that Equity\PnL lengths can be any length, and Wikipedia implies that you only measure per year and nothing else where a clear number of daily returnees.

Read http://economic-definition.com/Other_branches_of_mathematics/Koefficient_Sharpa_Sharpe_Ratio__eto.html for example.

Anyone who has done a lab at least once should know about SR rationing.

Standard Deviation of Profitability. This old friend of ours: we thought we had smashed him to smithereens, but no; here he is, rising from the ashes to participate as a risk component in the calculation of risk-adjusted returns. Note to yourself that it is extremely important here to express this statistical value for the appropriate time interval-ideally, as mentioned above, for one year. Because of the specific nature of this calculation (where this figure varies as a direct function of the square root of the number of partial values of the observations), this requires either multiplication or division of the square root of the number of observations. For example, suppose you have daily data for a year that determines a daily standard deviation of, say, $10,000, or 1% (let the amount of capital be $1 million). To find the annualized standard deviation, multiply that figure by the square root of the number of operating days in the year. If you cross off weekends and holidays on the calendar, you get about 250 plus or minus one or two days, and the square root of that number is about 15.9. Therefore, if the daily standard deviation is $10,000, or 1%, then the annualized standard deviation is about $159,000, or 15.9%.

In the formula for calculating the Sharpe Ratio, this normalization over time intervals must be done in order for the results to make sense. Note that this formula allows for adjustments for factors such as the fact that the data set may not be complete (e.g., six months of data), and that the time periods will not necessarily equal one day. However, I will rely on the opinions of my statistical professional friends to explain these puzzling phenomena.

At this point, you've probably already calculated your Sharpe ratio, and you're wondering whether you should be ashamed or proud of the result. Following a simple rule of thumb, I think you should almost always aim for the Sharpe ratio calculated using the above method to be greater than or equal to one. For example, assuming a risk-free interest rate of 5% and an annualized standard deviation of income of 15%, such a portfolio would need a return of at least 20% to reach this threshold:

0

[pantural]

Rashid Umarov:

No need to hide behind clever words. Here we take only PnL sample without any timeline and use it to calculate the Sharpe Ratio.

resample - as an obvious way to falsify the algorithm SR

Change the algorithm quickly, before someone notices))))

[Rashid Umarov]

pantural:

resample - as an obvious way to falsify the algorithm SR

change the algorithm quickly, before someone notices))))

Got it, thanks. To the ban

[Rashid Umarov]

Aleksey Vyazmikin:

I think you've found an explanation for why the Sharpe calculation slightly increases with the growth of the initial deposit. And by and large, the larger the initial deposit (basis for calculations) with the same absolute changes in balance/equity, the less volatility of funds in the account in relative values.

[Rashid Umarov]

pantural:

PS SR >3 is quite a normal SR value, unless of course it's a fit, with HFTs it can be double-digit:)

It's good that you know how to use search. https://smart-lab.ru/blog/267416.php

Most investors have probably never seen the equity of a high-frequency strategy. There are objective reasons for this: because of the typical performance of such strategies, firms using them have little need to raise outside capital. In addition, HFT algorithms have capacity constraints, which are very important for institutional investors. Therefore, it is interesting to observe an investor's reaction to the profitability of an HFT strategy he sees for the first time. Being used to a Sharpe ratio in the range of 0.5-1.5 or up to 1.8, under lucky circumstances he is amazed that such strategies show ratio values expressed in double digits.