Hypothesis formulation is the process of crafting a testable, falsifiable statement that attempts to explain an observation or predict an outcome. It is arguably the most creative step in science, yet it receives surprisingly little formal attention. Methodology courses focus overwhelmingly on how to test hypotheses, while the arguably harder question of how to generate them in the first place gets treated as a mysterious spark of insight that cannot be taught. That framing sells the process short, because researchers have identified dozens of concrete strategies for coming up with good hypotheses, and the way a hypothesis is framed shapes everything that follows in a study.
What a Hypothesis Actually Does
A hypothesis sits between an observation and an experiment. You notice something puzzling or interesting, you propose an explanation for it, and then you design a test that could prove the explanation wrong. That last part is critical. In the philosophy of science, Karl Popper argued in the 1930s that a statement only qualifies as scientific if it can, in principle, be shown to be false. A claim that cannot be disproven by any conceivable evidence is not a hypothesis in any meaningful sense; it is closer to a belief.
This falsifiability requirement has practical consequences. “People who sleep more are healthier” is too vague to be a real hypothesis because the terms are undefined and the claim cannot cleanly fail a test. “Adults who sleep fewer than six hours per night have higher resting cortisol levels than adults who sleep seven to eight hours” is testable: you can measure cortisol, compare the groups, and find out whether the prediction holds. The tighter and more specific the hypothesis, the more useful the test becomes.
A good research hypothesis shares several key features. It should be explicit in what it predicts, grounded in existing evidence, formulated before the experiment begins, and capable of being tested with real-world data. It should also carry explanatory power, meaning it does not just predict a pattern but offers a reason for it.1PubMed Central. Research Hypothesis: A Brief History, Central Role in Scientific Inquiry, and Characteristics Each of those qualities serves a different purpose. Being explicit prevents moving the goalposts after the data come in. Being evidence-based keeps the hypothesis grounded rather than speculative. Being stated before the experiment ensures the test is genuinely confirmatory.
Where Hypotheses Come From
The popular image is that hypotheses arrive in a flash of genius, but the reality is more systematic than that. One influential framework describes hypothesis generation as “abductive reasoning,” a process in which researchers take various clues, identify restrictions, and combine existing ideas in novel ways to propose a plausible explanation for something they have observed.2PubMed. Abductive reasoning and the formation of scientific knowledge within nursing research Abduction is not the same as deduction or induction. Deduction starts with a general rule and predicts a specific case. Induction starts with many specific cases and infers a general rule. Abduction starts with something surprising and works backward to find the best explanation. It is closer to detective work than to textbook logic.
Beyond abduction, psychologists have catalogued a surprisingly large toolkit for generating hypotheses. One review described 49 distinct creative heuristics that researchers have actually used in practice, organized into categories ranging from simple common-sense observation of oddities in everyday life to sophisticated ways of re-analyzing existing data to provoke new insights.3PubMed. Creative hypothesis generating in psychology: some useful heuristics The key finding was that every one of these heuristics is teachable. Hypothesis generation is a skill, not a talent, and treating it as unteachable shortchanges students who never learn how to think generatively about research questions.
Some of these heuristics are straightforward. Noticing that a commonly accepted claim does not match your own experience is one. Deliberately looking for exceptions to a well-established finding is another. Borrowing a framework from one field and applying it to a different domain has produced some of the most productive hypotheses in science. Other heuristics are more technical: running exploratory statistical analyses on large datasets and then treating the patterns that emerge as candidate hypotheses rather than confirmed findings. What they all share is that they are deliberate strategies, not accidents.
Searching Two Spaces at Once
One reason hypothesis formulation is hard is that it involves parallel searches in two separate domains: the space of possible explanations and the space of possible experiments. You are not just looking for a good idea; you are looking for a good idea that you can actually test with the resources and methods available to you. Research on how people develop scientific reasoning skills has shown that adults tend to coordinate these two searches simultaneously, going back and forth between “what could explain this” and “how would I test that,” refining both as they go.4PubMed. Heuristics for scientific experimentation: a developmental study This dual-search process is a domain-general skill, meaning it applies across fields and does not depend on specialized subject knowledge.
This has an important implication: someone trained in rigorous hypothesis formulation in one discipline can transfer that skill to another. A chemist who moves into biology does not start from scratch in learning how to generate good research questions. The content changes, but the cognitive architecture of searching for explanations and matching them to feasible experiments stays the same.
Exploratory Versus Confirmatory Research
Not all research begins with a fully formed hypothesis, and that is fine as long as the distinction is kept clear. Exploratory research is the stage where you are looking at data, noticing patterns, and generating candidate hypotheses. Confirmatory research is the stage where you take a specific pre-stated hypothesis and run a rigorous test of it. Both are legitimate and necessary, but mixing them up creates serious problems.5Significance. Different Worlds Confirmatory Versus Exploratory Research
The most common way they get confused is when a researcher explores data, finds an interesting pattern, and then writes up the paper as though the hypothesis existed all along. This makes the finding look much stronger than it actually is, because a hypothesis generated from the same data it is tested on has a built-in advantage. The data already fit the hypothesis by construction. In preclinical research, distinguishing clearly between exploratory work, where the goal is generating theories about how a disease works, and confirmatory work, where the goal is demonstrating strong and reproducible treatment effects, has been identified as critical for improving the rate at which lab findings translate to clinical success.6PubMed Central. Distinguishing between exploratory and confirmatory preclinical research will improve translation
The practical lesson is that exploratory analysis is a hypothesis-generating tool, not a hypothesis-testing tool. If you find something interesting in an exploratory phase, the next step is to write it down as a formal hypothesis and test it in new data. Skipping that second step is one of the most common sources of findings that fail to replicate.
Cognitive Traps in Hypothesis Formulation
Human minds are not naturally well calibrated for fair hypothesis testing, and the biases start during formulation itself. Confirmation bias, the tendency to seek out information that supports your existing belief while ignoring information that contradicts it, has been documented in experimental settings designed to mimic scientific reasoning. In one classic study, participants tasked with testing hypotheses in a simulated research environment showed a strong tendency to choose test environments that could only confirm their hypothesis rather than environments that could also rule out alternatives.7Quarterly Journal of Experimental Psychology. Confirmation Bias in a Simulated Research Environment: An Experimental Study of Scientific Inference
This is not about dishonesty. Confirmation bias operates unconsciously. Researchers genuinely believe they are being fair, but they frame hypotheses in ways that make their preferred explanation easy to confirm and frame alternative explanations in ways that make them easy to dismiss. The result is a research landscape tilted toward confirming what people already believe rather than discovering what is actually true.
A related problem is what happens when researchers formulate hypotheses after seeing their results and then present those post hoc explanations as though they were predictions made in advance. This practice has been given the memorable acronym HARKing, for “Hypothesizing After the Results are Known.” The concern is not just that it is misleading in any single study, but that it inflates the apparent success rate of predictions across an entire field, contributing to what has been called the replication crisis.8Review of General Psychology. When Does HARKing Hurt? Identifying When Different Types of Undisclosed Post Hoc Hypothesizing Harm Scientific Progress When every published study appears to have predicted its findings, readers lose the ability to tell which predictions were genuine and which were retrofitted.
Preregistration as a Structural Fix
One of the most significant recent reforms in how science handles hypothesis formulation is preregistration: publicly recording research questions and analysis plans before collecting or examining data. The logic is straightforward. If the hypothesis and analysis plan are documented in advance, there is no ambiguity about whether the researchers predicted the outcome or discovered it after the fact. Preregistration cleanly separates predictions from postdictions.9PubMed Central. The preregistration revolution
That said, preregistration is not a panacea, and the quality of preregistrations varies enormously. A study comparing structured preregistration templates, which walk researchers through specific fields to fill in, against unstructured formats found that structured templates did a better job of limiting researcher flexibility, but neither format eliminated all room for post-hoc adjustments. Perhaps more telling, when independent coders tried to identify how many hypotheses were stated in a given preregistration, they agreed only about 14 percent of the time, indicating that hypotheses were often not stated clearly enough for someone else to identify them unambiguously.10PLOS Biology. Ensuring the quality and specificity of preregistrations The lesson is that writing down a hypothesis is necessary but not sufficient. Writing it down with enough clarity that a stranger could verify what you predicted is the actual bar.
Hypotheses as Models
In many fields, hypotheses are not just verbal statements but mathematical models. A mechanistic model, which tries to represent the actual processes underlying a phenomenon rather than just fitting a convenient curve to data, is itself a hypothesis about how something works. Because the math is precise, the predictions are precise, which makes the comparison between prediction and observation much sharper than it would be with a verbal hypothesis.11PubMed. Mechanistic models are hypotheses: A perspective
This perspective has a practical consequence for how model parameters should be treated. If a model is genuinely a hypothesis about underlying processes, then its parameters should ideally come from independent measurements or first principles rather than being tuned to fit the same data the model is trying to explain. When you fit parameters to data, the model will look good by construction, similar to the problem with HARKing at the level of verbal hypotheses. When you derive parameters independently and then see whether the model’s predictions match observations, you are running a much more honest test.
Computer simulations take this a step further. By encoding a hypothesis as a simulation, researchers can watch how a complex system would behave if the hypothesis were true and compare that behavior to what actually happens. Simulation models allow researchers to test ideas before running expensive or time-consuming physical experiments, creating an iterative loop where the model is refined and the hypothesis sharpened before resources are committed to a definitive test.12PubMed. Computer simulation studies and the scientific method In fields like climate science, epidemiology, and engineering, this simulation-based approach to hypothesis formulation is now standard practice.
The Problem of Underdetermination
A challenge that haunts hypothesis formulation across all fields is that the relationship between theory and evidence is rarely one-to-one. Multiple different hypotheses can often explain the same set of observations equally well. This is sometimes called the Duhem-Quine problem, and it means that finding evidence consistent with your hypothesis does not necessarily rule out competing explanations. The same data might fit several very different stories about what is going on.
What is less often discussed is that the problem runs in both directions. Not only can multiple theories explain the same evidence, but the evidence itself is not always clear-cut. Describing what was actually observed in an experiment involves its own set of assumptions and interpretive choices. A measurement is not raw reality; it is reality filtered through instruments, calibration choices, and decisions about what counts as a data point versus noise. This means that researchers making different reasonable choices about how to characterize their data could end up with different empirical “facts” from the same experiment, each of which might support different hypotheses.
The practical response to underdetermination is not despair but discipline. It means formulating not just one hypothesis but also specifying the most plausible alternatives and designing experiments that discriminate between them. A study that can only confirm your preferred explanation, and cannot distinguish it from the next most likely alternative, is a weaker test than one that pits two hypotheses against each other. This is one of the reasons falsifiability matters so much at the formulation stage. A hypothesis that could be false, and that specifies the conditions under which it would be shown to be false, is far more useful than one that accommodates any possible result.
How Funding Shapes What Gets Hypothesized
Hypothesis formulation does not happen in a vacuum. The questions researchers choose to ask are shaped by what funding agencies are willing to pay for, and this introduces a form of bias that operates before any data are collected. Funding priorities set by government agencies and private foundations direct scientific attention toward specific applied outcomes, which can lead to the neglect of fundamental research and alternative ways of framing a problem. When a grant program defines the questions it wants answered, researchers who want funding naturally formulate hypotheses that fit within those boundaries.
This is not necessarily corrupt, but it does mean that the hypothesis landscape of any field is partly a map of its funding landscape. Questions that are easy to fund get tested. Questions that are hard to fund, even if they are scientifically important, may go unexplored for years. Reliance on specific funding sources can also create subtle conflicts of interest that influence not just which hypotheses are formulated but how experiments are designed, how data are interpreted, and which results get reported. Awareness of these pressures does not eliminate them, but it helps when evaluating why certain topics have been studied intensively and others have been largely ignored.
AI-Generated Hypotheses
One of the most active frontiers in hypothesis formulation involves artificial intelligence. Large language models and other AI systems are now being used to generate hypotheses at a scale and speed that no human researcher could match, by computationally recombining existing knowledge in novel ways.13ACS Materials Letters. AI-Generated Hypotheses and the Emergence of Autonomous Scientific Discovery The idea is not that AI replaces the creative act of hypothesis formulation but that it dramatically expands the search space, surfacing combinations and connections that a human expert might never consider simply because no one can hold the entire literature of a field in their head at once.
Recent work has shown that language models can generate hypotheses from labeled examples, iteratively improving them in a process inspired by how multi-armed bandit algorithms balance exploring new options with exploiting known good ones. On real-world classification tasks, the hypotheses generated by these systems improved predictive accuracy substantially compared to simpler approaches, and the generated hypotheses sometimes uncovered insights that were new even to domain experts.14arXiv. Hypothesis Generation with Large Language Models
The idea of automated hypothesis generation is not entirely new. Over two decades ago, a physically implemented robotic system demonstrated the ability to originate hypotheses about functional genomics, design experiments to test those hypotheses, run the experiments using a laboratory robot, interpret the results, and repeat the cycle autonomously.15Nature. Functional genomic hypothesis generation and experimentation by a robot scientist What has changed since then is the power of the language models and the breadth of knowledge they can draw on. Early robot scientists worked within narrow, well-defined domains. Current AI systems can range across entire literatures and propose cross-disciplinary connections that were previously invisible.
The coupling of AI-driven hypothesis generation with autonomous laboratory platforms raises the possibility of end-to-end automated discovery, where a system identifies gaps in knowledge, proposes explanations, tests them, and updates its understanding without human intervention at any step. Whether that changes the nature of scientific understanding or just accelerates it is a question philosophers of science are only beginning to grapple with. For working researchers, the more immediate question is practical: can AI-generated hypotheses be trusted enough to guide expensive experiments? The early evidence suggests they can generate plausible, testable ideas, but the filtering, prioritization, and ethical oversight still require human judgment.
A Skill Worth Practicing
If there is one underappreciated aspect of hypothesis formulation, it is that the quality of the hypothesis determines the ceiling of the entire study. A brilliant experimental design testing a poorly formulated hypothesis produces, at best, a clear answer to a question nobody needed answered. Conversely, a well-formulated hypothesis can survive a mediocre study design, because even a rough test of a genuinely important and precise prediction advances understanding. Researchers who invest time deliberately practicing hypothesis generation, using heuristics from cognitive psychology, learning from the structure of successful hypotheses in their field, and paying attention to the anomalies and surprises in their own data, are building the single most transferable skill in science. It is the one step that no amount of statistical power, funding, or laboratory equipment can substitute for.