DonFerrari said:
Not sure if you know metrology, but you can also decentralize your "mean"... instead of saying 10.0+-0.05 you could say 9.95+0.1. In statistic don't know it would be acceptable... but I never saw a statistic report on election being 22.5+-.5% of intention (and they use integer numbers even tough they state results in percentages), and then they give margin of error as 5 pp (percentual points)... by your logic they should state it as 22.5341% (if that was the exact approximation value on their method on lets say 1430 people polled). |
I know you can do it and yes, if you decentralize and you do like you say, that is taking into account the additional error you introduce doing so, then it's correct, but this way you are unnecessarily increasing the error margin, and unnecessarily worsening the error accumulation in the lifetime total. And yes, it also makes sense for immediateness to round to an integer percent value when you communicate a percent result, or to just one decimal, like ioi does in the bar chart in VGC's front page, but you don't want to do the same for data that you'll reuse for further calculations, like adding weekly absolute numbers to get monthly, yearly, LTD and lifetime totals. Then, when using those numbers to draw a curve, rounding those integers to the resolution available will again be correct, but it's correct to do it for FINAL data that you won't reuse. If you'll want to draw another graph, you'll start again from the unrounded values, you won't get them scanning the graph you previously made rounding the values. If you round too much and too early data that must be reused, the error on further results could skyrocket. BTW, rounding to the closest hundred or thousand can be a significant additional percent error for small weekly numbers.
Then there is a method issue: ioi devised his formulas for extrapolation and he needs to constantly refine, update and possibly improve them or at least not let them worsen as the market changes, to achieve it he cannot arbitrarily round numbers, he could do it for some samples if he found that a given source is systematically above or under the average and its error isn't random, but biased up or down, but ouside of these cases, he'll want to keep all his numbers as centered as possible, then he'll only round data that he just presents to the public without reusing them, like the percent values and rounded totals of the front page bar charts. Another thing: real sales numbers are integers, so after applying extrapolation formulas, that will mostly give results with decimals, rounding to the closest integer is right and necessary. But after doing it, an integer like 370022 has the same "dignity" as 370000, rounding it to the closest hundred or thousand presenting the results is just cosmetic, you'll do it regularly if you, for example, present the data as 370k, 0.37M or 0.4M and so on, but again, you won't add these quick representations of weekly totals to calculate yearly totals. The 0.37M example can give you a hint: if you write, for example, 10.1M for a lifetime total, your rounding error is at most ±0.5%, but if you decide to use just ONE decimal and you write 0.1M for a weekly total that before cosmetic rounding was a given number somewhere between 50000 and 150000, your rounding error can be up to 50%: even when deciding to do cosmetic roundings, you'll have to consider the range of actual values before deciding the acceptable rounding.
EDIT: I completed this post after taking some time and a long pause and in the meantime ioi wrote another comment on this issue, read it too, as he's enormously more skilled than me in the statistics and probability field (I, despite not liking them very much, except for robotics, found that I'm better at automatic controls, and, for example, I have many time had the temptation to represent ioi's adjustments as a linear system with a feedback loop ).