Hacker News Reader

Comment by dv_dt | Hacker News Reader

Despite the text implying it's all an exponential curve forecast model, the wiggles seem to point to actual data being plotted too.

I wish they differentiated data points from forecast points in the actual chart. Text says 1/29 update, but I think only the 1/28 point seems to match news reports.

View on HN

It could be clearer, but, the way I read it was that it's the historical data plotted + 3 days of predictions.

Which is 100% misleading, and strips all accountability from the predictor. Show us your prediction and the difference between the actual and predicted results to see how well you did.

Stop moving the goalposts too. Make the prediction for all days in the future once, and don't change it based on the current numbers.

https://news.ycombinator.com/item?id=22185174

Though near-term projection based on current trend is useful.

Eventually more humans will die from it than have ever been born! ;)

It's the Four Yorkshiremen Epidemic.

It's just an exponential fit based on the current number. You can run that same fit for earlier days and eyeball the graph, but we have a fancier tool for that called a "correlation coefficient".

> Stop moving the goalposts too. Make the prediction for all days in the future once, and don't change it based on the current numbers.

So, don't ever improve the prediction based on new data. Mkay.

Extrapolating out into the future is always errorprone. Exponents don't ever continue forever and they look like great models until they don't.

You can definitely improve your prediction based on new data, but the most important part is to show your historical predictions too. FiveThirtyEight does this pretty well by showing the probability a candidate gets elected in a time series chart.

Predicting is serious business, and you have to be accountable for past predictions, especially if they were wrong.

This chart, as is is extremely misleading because it gives readers the false sense that you were 100% right, and your future predictions are highly likely to be true.

>> Extrapolating out into the future is always errorprone.

You're absolutely right. And if readers saw your predictions were mostly wrong, they wouldn't be alarmed in the future when they see "experts" predict 10% of the world would die 3 months after a virus is found. Because they'd know from past experience that past models were mostly wrong!

> This chart, as is is extremely misleading because it gives readers the false sense that you were 100% right, and your future predictions are highly likely to be true. Instead, if readers saw your past predictions were mostly wrong, they wouldn't have this false perception.

Just hit the button to toggle to the log scale version of it. A exponential fit is appropriate and past predictions using an exponential fit would be good. :P Do you understand how to fit data series? My fifth grader could hold up a ruler-- align it the the beginning of the yellow line and each subsequent point and see how reasonable of a prediction they are of future points. He'd say "Yup, it's pretty straight!" And even for the red line, though a little less so.

Yah, it would be cool if they'd tell us r^2.

> And if readers saw your predictions were mostly wrong, they wouldn't be alarmed in the future when they see "experts" predict 10% of the world would die 3 months after a virus is found.

Note that exponential models predict everyone will get the disease and everyone will die if you extrapolate out far enough. (Actually, -more- than everyone heh). Obviously we don't think that's the case. But it's a pretty damn reasonable way to see 3 days forward in the early phases of an epidemic.

agreed. I mention there that this is simplistic. I may look into the sigmoid shape s-curve once it is time for it. If the site is still interesting to somebody at that time.

Yes you are right. I moved the data one day forward as it seems more natural to call data that we received today as todays data and not yesterday's, but that was a bad idea. I fixed that.

Also I tried to add as much info about the sources I had now. I will try to improve on that later. The site mentioned here https://news.sina.cn/zt_d/yiqing0121 really looks like finally one good source of the data I was looking for.

I need to get some sleep now.

Mostly I was hoping for a different color and/or symbol for data vs prediction points as a way to more quickly understand what the chart curve represented.

Another interesting comparison would be to draw a model curve (or curves) further back - the difference vs data points might give you a visual indicator of how the virus is spreading vs effect of active/natural influences vs the model.

Thanks for the update, don't be too hard on yourself, there are infinite ways to improve, but I appreciate the quick look data.

I colored the prediction points in blue.

It seems that actual are integer values while predictions are float numbers.

And cited sources for their data. And had error bars on the prediction. And gave you access to historical predictions.

Historical case counts don't really have error bars... and it's difficult to come up with meaningful error bars based on goodness of fit of a function.

But the graph is also looking three days into the future, presumably based on fitting some model to the historical data. The output of any worthwhile model is a probability distribution over possible future infection and death counts. As a consumer of that graph, I should be able to see not just the value just a model thinks is most likely, but at least some indication of how tight a bound the model is placing on those values.

The lack of which suggests that the author (and who is that, exactly?) may not have given a moment's thought to any kind of responsibility they might have in the midst of a health crisis.

Oh, what the hell. Let's write some Javascript! And slap a tweet button on it!

If you just want the historical data, try https://news.sina.cn/zt_d/yiqing0121 Of course the labels and data sources are all in Chinese...

Reply on news.ycombinator.com