My very first question here has been closed as "unclear what you're asking".
The "How to Ask" section says
- Is your question about statistics?
As far as predictions belong to statistics.
- We prefer questions that can be answered, not just discussed.
And so do I.
- Provide details. Share your research.
I posted everything I could. I could do more, but with little effect:
- I could invent some ad doc data-clustering algorithm, which might fit the bill, but I'd prefer some more scientific approach - where I have nothing to say.
- I could implement some machine learning, but that seems to be an overkill for this task.
The task at hand is really specified as I wrote: Do some prediction, it's clear that it can't be perfect.
I surely agree that my question could use some more details and I'll gladly add them. However, I disagree that it can't be answered.
Reply to comments by Glen_b
I think your question still seems pretty vague about what precisely is required
This is no research task with some error function to be minimized. There's a customer's wish to "make some predictions", so I'm trying. I'm sure, I'm not the only one and actually, I'd expect this to be well-known problem. I'm well aware that given the tiny and irregular data sets, not much can be done. OTOH something can be done and the quality of the result can be evaluated on the current and future data.
it talks about how something "seems"
I was deliberately using such phrases as I'm only describing how I think the output for a given data set should look like. A different output is acceptable.
guess how that might generalize for your feelings about how things seem
You don't need to generalize my feelings, just provide any solution. Or multiple solutions. Imagine, I was asking about how the sequence
1, 2, 3 continues. You could tell me, that it's impossible to tell, and it'd be correct. Correct but useless. Or you could point me to the arithmetic progression and Fibonacci series and whatever and I could try it and verify the predictions using future data.
an answer would rely too heavily on opinion about how to interpret it
So don't be scared and give me some answer. It may fit and then we're done. Or it may not, and then I'll see what I've missed.
If there wasn't the problem with events around midnight, I'd simply determine the most common weekday (using some decay) and compute the average time and then look how good it works. I guess, this alone would allow to predict the next event with some 80% certainty and two hours tolerance in half the data sets (some data sets contain events happening bi-weekly, etc.) and even this could be an acceptable solution for now.
Reply to the answer by Scortchi
You seem to be part-way through developing a heuristic method for prediction
Actually, I still haven't really started besides extracting the input data and analyzing maybe six data sets manually.
it's concerning that my knee-jerk prediction for the first example is quite different from yours
Yours is equally valid. In fact, it's something I haven't though about. And the data continues: 180320-Tue-2222. So you were right.
the importance of providing context
IIRC you're the first telling me about it. I simply left the context out, as I don't think it's important for this question (though it surely is for many others). I may be right as you lacking context could make a better prediction than me.
Can you share the data?
I guess so, I'll ask.
(tons of question)
I'll answer them in the question.
leading to a mare's nest of conflicting answers
I really don't think, it'd be a problem. I may be wrong...
I'll reword the answer, provide the context and add lots of data. We'll see if it helps.