Alright, I am in a bind.

I know that prediction is the ultimate test of a theory in the physical world. You can make any assumptions, you can come up with useless things like the math in string theory, but ultimately it's only considered "true", as true as such theories could be if and when they make a testable prediction, and it comes out to be accurate.

Where do we draw the line on the accuracy of the prediction before we call it a true theory or model of the real world? Predictions can vary in accuracy, for eg: Sun rises in the east, vs. sun will rise tomorrow at this and that angle at this time and so on.. And then you have to consider the frequency of the data or pattern that you observe as well into consideration. Now I know that if the pattern is not seen often, then you'd want to make a far more accurate prediction to prove the theory true vs. when a pattern repeats often, is observed in multiple places, because then in the latter case, since you can make multiple predictions you could be lax with accuracy and make more general statements, for eg: things with mass are attracted to each other vs. things with mass at a distance, have a mutual attraction represented by a quantity called "force" which is inversely proportional to the square of the distance between them and their masses and so on bla bla.

Do we always need an idea based approach first? Since it has the power of prediction compared to a more data fitting based approach ( which might not have any underlying intuitive reasoning)? I like the idea based approach because there is less "peeking" at the data, and more coming up with sound reasoning.

Lastly falsifiability. This thing has been bothering me a little. I thought it was central to calling a physical theory of the world, a "theory". You need to have a concept of falsification otherwise it is in the same bucket as religion, but doesn't making a prediction encompass falsifiability? It seems to me that falsification is a test of validity and not truth. Am I correct in this? Because if you have a statement which says: if A then Y, then you automatically have the falsification criteria satisfied.

  • Franz answer does show me that I am not way off. Please feel free to add your viewpoints as well. I want to think of this from every goddamn possible angle.
    – user12196
    Jul 23, 2017 at 6:37

6 Answers 6


It's not so much the predictive success of theories that convinces us that they are true than the fact that they make novel unexpected predictions when they are extended to new area and when there's a convergence between different experimental area.

Take for example Newtonian physics. It failed to account for the trajectory of Mercury. It wasn't considered false for this reason, rather scientists postulated hidden planets or asteroids near the sun to account for this. If a theory fails in its predictions, it is always possible in principle to make additional hypothesis that will restore its truth. But what we're expecting is that at some point, these additional hypothesis will be confirmed by independent observations: they must make novel predictions, not only account for what we already observed.

Now take the theory of relativity. It can account for the trajectory of Mercury. But that's not what convinced scientists that it was true: it was really confirmed when we tested a novel unexpected prediction of the theory, that light is deflected by massive objects. This was observed during a solar eclipse (the apparent position of stars was altered by the presence of the sun) and that was the decisive experiment that led to a wide acceptance of this theory by the community.

So I wouldn't say that it's a question of level of accuracy (because we can always postulate noise or measurement errors to account for them), but rather the fact that there's a conjunction between very different area of experience and the theory unifies all these phenomena in a single simple framework that explains them all, and that it even continues to work when applied to new areas. That's what is decisive for theory acceptance, not a certain level of accuracy.

I think this is the sense following which we need an "idea based approach" as you say: a data fitting approach is incapable of unifying different area of experience in a simple explanatory way. Or to say it differently, science postulates unobservable things to explain observable phenomena (and unify them).

As for falsification, it is indeed a criteria of scientificity, not of truth, but this criteria can be criticised, precisely for the reason I gave: it's always possible to "save" a theory with additional hypothesis, so theories are actually not falsifiable (strictly speaking) and scientists never reject well accepted theories at the first experimental failure.

  • Makes sense. Wow. I could make such a prediction.
    – user12196
    Jul 24, 2017 at 13:45
  • Thank you, it's a new angle I did not think of before. This seems like something which would avoid "over fitting".
    – user12196
    Jul 24, 2017 at 14:46
  • 1
    That's related to the debate on scientific realism, and what people call the "no miracle argument". Jul 24, 2017 at 17:17
  • sounds really interesting.
    – user12196
    Jul 24, 2017 at 20:08
  • 1
    Yeah, exactly. Over fitting sums it up.
    – user12196
    Jul 24, 2017 at 20:36

I agree this is an interesting question. I think you were wise to put the word "true" in quotation marks up there. If you look at the history of science, it seems that the truth of the future very, very often exceeds the truth of the present. So we can say that, in general, truth exceeds our scientific method of the time (truth exceeds method). (Gadamer). Many of our truths are therefore useful, but really just provisional, until the next "great one" comes along, etc. W.V. Quine, philosopher-logician had an interesting idea with his "Web of Belief" which can be read online I think. Don't know if there is a download or not.

  • Sounds goos man, and all true words. But does not solve my predicament. Strange no upvotes...
    – user12196
    Jul 24, 2017 at 4:38
  • I'll be more specific, if a legitimate journal accepts the work for publication, that's a least a good sign it may be true, or at least that it's interesting. Ultimately it's "true" when the community of scientists says it's true. Einstein published his paper destroying absolute space and time, nothing happens, months later he he gets a letter from Max Planck, months pass again... As far as ideas vs data, a scientist will get his start however he can, ideas, feelings, hunches, accidents...all good, then he goes to work.
    – Gordon
    Jul 24, 2017 at 5:48
  • This is very traditional, but Truth is truth, and does not depend on popular opinion. The sun will shine if I look at it or not. I did not ask for scientific theory popularity criteria sorry. I am sure even bull shit like String theory is published some where, in some very respectable journal.
    – user12196
    Jul 24, 2017 at 8:42

As far as physics is concerned, accuracy varies with the matter at hand. Think particle physics vs Newtonian physics. I don't think there is a universally acceptable definition or some margin error in percentage terms that is universally applicable to all fields. All observations have a measurement limit and no theory can be absolute "truth" in the sense it can never be verified with 100% accuracy, ergo questioning the very existence of a "true" theory. Of course for practical purposes it is inconsequential.

Sure the idea based approach sounds elegant as it appears to be based on physical foundations but essentially there is no foundation. All basic theories are just observations (i.e data) described neatly and extended as ideas in other areas. So is there really a dichotomy? Take gravitation, would you call it an idea or simply a theory based on data? or for e.g. the postulate that speed of light is a universal constant, an idea on which relativity stands. It was basically an attempt to fit data (or observations). The key, however, lies in using data from one field and using them to predict altogether novel ideas (e.g. gravitational waves) and then verifying them. But it all starts with "data fitting" not the other way round.

I agree that falsifiability is more about the approach to describing truth rather than the truth itself. But does Prediction always encompass falsifiability ? If X then Y - to falsify this, y has to be verifiable by experiment or observation, at-least in Physical Sciences. If its not verifiable it is certainly not falsifiable but will it be considered a prediction ?

  • You are wrong about the falsifiability part i think.
    – user12196
    Jul 25, 2017 at 14:00

It's an interesting question. The prediction level is probably explicitly defined in certain circumstances - e.g. discovering a new particle in the Large Hadron Collider like the Higgs boson might only be accepted if the statistical tests are significant (I don't know exactly what the case was for the Higgs boson but you get the point). The interesting thing is that for a lot of other theories it's not explicit. When the results of the experiment don't match the predictions given by a theory, there are two options - one is to say the theory is wrong/insufficient. The other way out is to say your measurement is error or the machines are faulty. E.g. none of the chemistry experiments we tried out in school ever worked out perfectly ,we were always off by a couple of mls (from what was expected) so we always assumed that we made slight measurement errors, not that the theories are wrong. I think a lot of this is based around what is practical - perhaps the universal constants aren't constant, and differ slightly, or perhaps the measurement noise is error. It depends on which theory makes more sense at the time, and whether or not the predictions of each continue to work the next ten times you test them.

Typically in science it is never purely idea based as you put it. You need some data to work with to create the predictive model (perhaps you may find some philosophers who disagree with this and believe it is possible to have a priori scientific knowledge). Peeking at the data includes making any observation surely (like noticing the sun set in a certain direction today)?

Lastly, I think you're right about falsification in that prediction gives you a way to show a theory was false. Ie if it predicts things in a wrong way, then you can show it doesn't work. The difficult part is when people talk about "explanations" that claim to be scientific but do not make predictions, e.g. a psychoanalyst describing the discovery of unconscious thoughts which are inaccessible to everyone but influence our every decision. It's not a good scientific theory as it doesn't make predictions but attempts to explain something that can't be shown to be true or false.

  • I am glad I do not have to worry about falsification. Making multiple predictions has a name in my dictionary - called consistency, lends a lot of value. So predictions, consistency and falsifiability huh. It will be hard to give a falsifiability argument in this case, and I think getting the consistency part could be hard as well.
    – user12196
    Jul 23, 2017 at 6:35
  • But, I would definitely need to reproduce my results right. Under the same circumstances if it yields the same exact result then the theory is absolutely correct and I am not missing out any factor? Is that approach correct? Must be.. Finding hard to argue vs. that. Please confirm though if you can. Thanks!
    – user12196
    Jul 23, 2017 at 6:48
  • Plus, yeah it should be extremely difficult to get absolutely non data peeking priori. Some peeking is of course required, but it is generally minimal, or prediction quality is good right?
    – user12196
    Jul 23, 2017 at 6:56
  • It depends on what you mean by your theory being absolute correct. I could flip a coin and see heads, and make a theory that the coin will always land heads. If I get heads the next time I flip, this doesn't guarantee I'm right. As another example, Newtonian mechanics makes great predictions for things our size, but doesn't work well for objects at speeds close to light speed or at the subatomic and cosmic scales. I would happily say Newtonian mechanics aren't wrong though in the sense that for many purposes they make great predictions, and that is science to me.
    – Franz
    Jul 23, 2017 at 7:03
  • I think you are wrong there. All physical theories are not guaranteed to be right whether it be QM, or relativity or tossing coins. You are only giving the coin toss example because you know it a priori.
    – user12196
    Jul 23, 2017 at 7:06

To some degree, I would disagree with what you say about prediction. I say this because using the word prediction could be misleading, when what I think we care about is building a theory on an accurate model of the phenomena, which gives an adequately descriptive account of the phenomena. And one theory would succeed another when it is seen to give a more adequately descriptive account of the phenomena.

  • This answer comes off more so as a comment than as an actual answer to the question. Do you think you could flesh your ideas out more, or maybe try to find some sources to site that argue for what you're arguing? Even still, it seems like this is just a comment and not really an answer. But maybe there's more there and expanding on the ideas will make it clearer.
    – Not_Here
    Jul 23, 2017 at 8:32
  • I disagree. Why do we need to describe the phenomenon? But i am curious as to your exact argument if you can elaborate. It is also impossible to put in description as an objective factor.( Describing description). PCA or neural nets are not descrptive. Work fine?
    – user12196
    Jul 23, 2017 at 10:57
  • Of course description is desirable. But how can it ever be a test?
    – user12196
    Jul 23, 2017 at 10:59
  • I guess what I am saying is we have an ideas based approach "all the way through". I do not mean that we should think of a theory as nothing more than a description of the phenomena, but I do think the base of a theory is its model of what is going on, its description of what is going on "behind" the phenomena - and when that model can no longer account for the phenomena/phenomenon, as it is observed, that is when the theory needs to be revised or replaced, since its model has in this case been shown to be not adequately descriptive in its account of the phenomena.
    – l_ruth_
    Jul 23, 2017 at 21:03
  • 1
    Interesting idea, I think I used to agree about science showing us the true nature of things, like describing things as they really are, but I think that ends up being much less coherent than the prediction based philosophy of science, because the prediction based philosophy does not require so much metaphysical baggage
    – Franz
    Jul 24, 2017 at 7:15

From a Kuhnian point of view, never. We choose an underlying theory that gives us a compelling story, and we hold onto that paradigm as long as it keeps netting us better and better results.

But we know from history that this paradigm will be displaced. We have yet to see a theory that just stayed the same, and was not either controverted or altered extremely at some point. If any of it was ever really true, that should not happen.

We can know how reliable a rule is, and even if we have a better rule, we can keep relying upon it to continue being as reliable as history proves it has been. It is as true as it is, it has a given statistical usefulness. We can know what theory is better, but we also know there is going to be a better theory yet, or science would just be in the process of digging its own grave.

You must log in to answer this question.