1

Topic: Algorithm of an estimation of coincidence 2

I welcome! Saw in this branch a subject "algorithm of an estimation of coincidence", I look, answer! Cool! I Will try to ask too! The task: 1) In an advertizing campaign there is a Conversion index. It in this case (number of orders) / (Number of shows of banners) = 0.1 or 10 % Advertising was shown N days. N - it is not known!! We tell in 1 day (either hour, or minute, or second and ) 100 shows and 10 orders, and can and 117 shows and 0 orders or 30 shows and 30 orders. Conversion of 10 % this averaging for all time of show of advertizing. The dispersion is not set and unknown and not . Access to the day statistics is not present. Only to total - all N shows, To - orders, To / N = 0.1 or 10 % however N not small, i.e. such that most likely in  an average of 0.1 or 10 % allocation is close to normal, but it is not known. On statements of the problem it is not known, but only my assumptions (like reasonable!) 2) the Manager made certain changes to adjustments of advertizing which could affect conversion. 3) a question: to Find minimum 1 and N1 such, what with probability of X % we can define that Conversion  on 1 % i.e. there were> =11 % 4) 2, N2 Y % - Conversion <= 9 % 5) 3, N3 Z % - Conversion did not change almost, i.e. 9 % <To <11 % test value X = Y = Z = 95 % still figure In statements of the problem a certain parameter Capacity [samplings?] 80 % the Dispersion is not known, but it did not change after advertizing campaign modification. About confidential intervals read, but did not understand as them to apply to the answer on 3) 4) 5) P.S. In the Internet a certain formula for calculation N - however its substantiation and an output is not present, so its value zero, the decision should be justified. Results in it turn out too big, the order of 10 thousand conversion and 100. Shows. Intuitively it should seems difference to be visible on the smaller statistics? And  to depend on a dispersion or not?

2

Re: Algorithm of an estimation of coincidence 2

Hello, loginx, you wrote: L> P.S. In the Internet a certain formula for calculation N - however its substantiation and an output is not present, so its value zero, the decision should be justified. L> results in it turn out too big, the order of 10 thousand conversion and 100. Shows. Intuitively it should seems difference to be visible on the smaller statistics? L> and  to depend on a dispersion or not? Indeed if you type K value from N events that an error there will be the order ~sqrt (K), a fractional error err=1/sqrt (K) K ~ err ^-2 = 0.01 ^-2 = 10 000 N ~ K/10 % = 100 000 5 % => K ~ 400 and N ~ 4000 10 % => K ~ 100 and N ~ 1000

3

Re: Algorithm of an estimation of coincidence 2

Hello, loginx, you wrote: L> I Welcome! L> saw in this branch a subject "algorithm of an estimation of coincidence", I look, answer! Cool! L> I Will try to ask too! L> the Task: L> 1) In an advertizing campaign there is a Conversion index. It in this case (number of orders) / (Number of shows of banners) = 0.1 or 10 % L> Advertising was shown N days. N - it is not known!! L> we tell in 1 day (either hour, or minute, or second and ) 100 shows and 10 orders, and can and 117 shows and 0 orders or 30 shows and 30 orders. L> conversion of 10 % this averaging for all time of show of advertizing. L> the dispersion is not set and unknown and not . L> Access to the day statistics is not present. L> Only to total - all N shows, To - orders, To / N = 0.1 or 10 % L> however N not small, i.e. such that most likely in  an average of 0.1 or 10 % allocation is close to normal, L> but it is not known. On statements of the problem it is not known, but only my assumptions (like reasonable!) L> 2) the manager made certain changes to adjustments of advertizing which could affect conversion. L> 3) a question: to Find minimum 1 and N1 such, what with probability of X % we can define that Conversion  on 1 % i.e. there were> =11 % L> 4) 2, N2 Y % - Conversion <= 9 % L> 5) 3, N3 Z % - Conversion did not change almost, i.e. 9 % <To <11 % L> test value X = Y = Z = 95 % L> still figure In statements of the problem a certain parameter Capacity [samplings?] 80 % L> the Dispersion is not known, but it did not change after advertizing campaign modification. L> about confidential intervals read, but did not understand as them to apply to the answer on 3) 4) 5) L> P.S. In the Internet a certain formula for calculation N - however its substantiation and an output is not present, so its value zero, the decision should be justified. L> results in it turn out too big, the order of 10 thousand conversion and 100. Shows. Intuitively it should seems difference to be visible on the smaller statistics? L> and  to depend on a dispersion or not? In the theory the dispersion can be most . For example: at first 1000000 shows of a banner, and after 1000 orders. Here it is better to look from the practical point of view. If to construct the schedule where on one axis an amount of shows, and on another - average conversion from the observation beginning on the schedule conversion at a certain amount of shows almost ceases to change. Here also it is necessary to take this amount of shows for a basis. To make algorithm which calculates conversion in such a way it is possible.

4

Re: Algorithm of an estimation of coincidence 2

Hello, Qulac, you wrote: Q> In the theory the dispersion can be most . For example: at first 1000000 shows of a banner, and after 1000 orders. Here it is better to look from the practical point of view. If to construct the schedule where on one axis an amount of shows, and on another - average conversion from the observation beginning on the schedule conversion at a certain amount of shows almost ceases to change. Here also it is necessary to take this amount of shows for a basis. To make algorithm which calculates conversion in such a way it is possible. Yes, but the task in in advance to predict it "on one axis an amount of shows" That is to name this number without fulfilling anything! With  probability... I.e. how many it is necessary to make conversion that with the given probability to state that it changed After all there can be 200 million, and there is no person in the Russian Federation so much. I thought so - it is possible to calculate confidential intervals for the first and second variant +/-1 % If they are not intersected that -  certainly better 94 % And if will be intersected?

5

Re: Algorithm of an estimation of coincidence 2

Hello, loginx, you wrote: L> Hello, Qulac, you wrote: Q>> In the theory the dispersion can be most . For example: at first 1000000 shows of a banner, and after 1000 orders. Here it is better to look from the practical point of view. If to construct the schedule where on one axis an amount of shows, and on another - average conversion from the observation beginning on the schedule conversion at a certain amount of shows almost ceases to change. Here also it is necessary to take this amount of shows for a basis. To make algorithm which calculates conversion in such a way it is possible. L> Yes, but the task in in advance to predict it "on one axis an amount of shows" L> That is to name this number without fulfilling anything! L> with  probability... What is the conversion? It is probability of that after show  the user makes the order (it was ). I.e. the task is reduced to measurement of probability of event on statistican. L> I.e. How many it is necessary to make conversion that with the given probability to state that it changed L> After all there can be 200 million, and there is no person in the Russian Federation so much. If real conversion 1 on 200 million that is required many shows for its more or less exact measurement.

6

Re: Algorithm of an estimation of coincidence 2

Hello, Qulac, you wrote: Q> That such conversion? It is probability of that after show  the user makes the order (it was ). I.e. the task is reduced to measurement of probability of event on statistican. Not measurement, and determination with probability of 95 % that this index changed more than on 1 % I found the Confidential Interval in the textbook, but it shows not change, and probability  a range of the given number. It almost it... But... Under the task at us will be 3 confidential  1)  initial number of conversion 2)  an interval from +1 % to 100 % 3)  an interval from k-1 % to 0 e-e-e-e-e as though to count an aggregate 2) 3) and a certain total confidential interval for it and probability...? And to calculate N or To necessary that probability  a composite interval was with probability of 95 % but similar there are no here people remembering a floor-mat. To statistican it is better than me...