I have been working on a research report in which forecasting is done on the basis of data, followed by an interpretation of the forecasted results.

Is it possible to have that kind of research without hypothesizing any statement?

If this question is off-topic kindly recommend a suitable community.

  • Would you please provide context for your thesis. For example, what is your area of study and what subfield are you in? – Richard Erickson Jun 14 '17 at 17:50
  • It would be answerable, if you could please add some more details as suggsted by @RichardErickson. – Coder Jun 14 '17 at 18:02
  • 3
    This question is not about academia but about statistics. It might be on-topic on crossvalidated.se – henning -- reinstate Monica Jun 15 '17 at 5:58
  • 2
    @henning I made it about statistics, but the question as it stands is about research protocol. On crossvalidated Ibn e Ashiq can ask how to do the statistics. – Joris Meys Jun 15 '17 at 8:52

As a statistician, I'm inclined to say "no, you can't" as a short answer. Reason for this is simple: even in complete random datasets on average 5% of the correlations will be significant when tested in a model. So if you rely on only the data to make any kind of interpretation on association of variables, you're bound to publish false positives. This has been discussed for decades, eg this rather strong opinion of Ioannidis (2005) :


He says:

Bias can entail manipulation in the analysis or reporting of findings. Selective or distorted reporting is a typical form of such bias.

If you don't formulate a hypothesis and still interprete the data, you're -probably unintentionally- reporting selectively. You select from the analysis those results that tell a story, and in doing so, you're likely to report something that isn't a solid association or relation.

That said, you don't always have to formulate a specific hypothesis. For example, if you compare multiple methods on efficiency, you don't have to hypothesize beforehand which one is going to be the best. But the statistical test you use for comparison, will imply a "null hypothesis" that there is no real difference between all methods. Also this is "formulating a hypothesis" merely by the choice of analysis tools.

And this is even more important to realize: you might not formulate a hypothesis explicitly, but the nature of the statistical tools you use to come to your interpretation, will imply a set of rather rigid hypotheses and assumptions. You need to be aware of those hypotheses and also of those assumptions.

Because that's something I see far too often: people not explicitly formulating a hypothesis, still interpreting results from a statistical methodology, but failing to realize that their data does not meet the assumptions of that methodology. And that invalidates your entire interpretation.

This problem is even more stringent when forecasting. If you use regression models, you should be aware that predictions outside the boundaries of the original data cannot be interpreted. The uncertainty on those predictions is simply too big. If you use spline methods, you can even get into trouble at the edge of your original data. So definitely in the case of forecasting I would write out both the goal of the research and what you expect the predictions to show, including the scientific reason why. Only in those cases you can use forecasts as some form of evidence for or against the expected relation. If you don't do that, your forecasting model might as well be a fancy random number generator.

So in conclusion: even if your research goal isn't necessarily a defined hypothesis, you still need to formulate the hypotheses you want to test before carrying out the actual statistical tests.

And in all honesty, writing down what you expect to see is always a good idea, even if it's only to order your own thoughts.

| improve this answer | |
  • 1
    +1 This is a really important point and a core principle of the scientific method, but for some reason many people either ignore it (out of convenience) or do not know about it. – 101010111100 Jun 14 '17 at 18:38
  • There are fields - especially in Engineering - where results do not depend on a statistical analysis so as non-statistician, I would say, yes you can write a thesis without a hypothesis. – o4tlulz Jun 15 '17 at 4:04
  • 1
    @Kevin That's using cross-validation, an often used statistical technique with its own assumptions. If you do that, you have to keep in mind that your training data and testing data have to be completely independent, or any hypothesis testing (yes, also there you test a hypothesis) is invalid. Independence is one of the most important assumptions in about every common statistical technique. – Joris Meys Jun 15 '17 at 8:47
  • 1
    @Ooker I always tell my students to only use methods they know and understand. You can't know everything about statistics. But what you need to know, is all the details of the techniques you use yourself. And in any case every student should have the knowledge of the basic tests used in the majority of papers. Because if you don't understand those, there's no way you can evaluate yourself whether the conclusion in a paper actually makes sense. – Joris Meys Jun 15 '17 at 8:51
  • 1
    @o4tlulz and how do you assure me that it solves the issue? Can you prove that? Can you show me your "solution" isn't just random luck? – Joris Meys Jun 16 '17 at 7:45

Not the answer you're looking for? Browse other questions tagged or ask your own question.