Why the quality ceiling of a prediction network is set by contributor supply, not by compute or data.
When we talk about the quality ceiling of a prediction network, the conversation gravitates toward inputs that are easy to count: compute capacity, data volume. We have come to believe that neither of those is the binding constraint. The binding constraint is the diversity and accessibility of contributor supply, because an aggregate is only ever as good as the independent views feeding it, and that independence is the genuinely scarce resource. This is not a theoretical preference; it is a conclusion we arrived at by watching what actually limits the quality of aggregated inference over repeated epochs and across shifting market conditions.
Consider the economics of each input. Compute is rentable at the margin and has been commoditizing for years; any participant who needs more can buy it, and the price continues to fall. The trend is structural: cloud providers compete on price, GPU availability expands with each generation, and inference costs drop on a curve that shows no sign of flattening. Historical data, particularly the price series, order-book snapshots, and on-chain records that short-horizon crypto forecasting depends on, is largely shared across everyone bidding on the same topic. Data vendors sell the same feeds to all subscribers, and the on-chain record is public by construction. When every participant can acquire the same inputs on similar terms, those inputs do not differentiate the output. They are table stakes rather than sources of edge.
What remains scarce is a genuinely different model: a contributor whose errors are not already reproduced by someone else on the same topic. That kind of input cannot be rented, and no spot market exists for it. In our experience, the value of a new contributor to an aggregate is almost entirely a function of how uncorrelated their errors are with those of the existing population. Two models that are individually accurate but wrong in the same places contribute roughly the same information as one of them alone. The marginal value comes from a model that is wrong in new places, because the aggregate can cancel errors that are spread across independent sources but cannot cancel errors that are shared.
The scarcity has a shape that makes accessibility as important as diversity. The people most likely to carry a decorrelated view are distributed across disciplines and institutions that have nothing to do with crypto infrastructure. A forecaster with a signal-processing background, a researcher carrying a prior from epidemiology or climate science, a quant whose edge sits in an unusual feature set. Their models are valuable to the network precisely because those models were built under assumptions the existing population does not share. We have observed this pattern repeatedly: the contributors who move the aggregate most are often those whose intellectual heritage is furthest from the median participant's. A climate scientist applying regime-detection methods developed for atmospheric data brings a fundamentally different error signature than a crypto-native quant trained on the same order-book features everyone else uses.
If contributing requires fluency in wallets, gas, epoch mechanics, and on-chain submission rather than in forecasting, the participants who clear that bar are selected for infrastructure tolerance. Infrastructure tolerance has no relationship to whether a model carries a useful, uncorrelated signal. The friction filters out exactly the supply that would have moved the aggregate. We have seen this in practice: talented researchers evaluate the onboarding path, estimate the hours required to learn protocol mechanics that have nothing to do with their craft, and quietly move on to something else. Each departure is invisible in the metrics, because the network never saw what it lost. The cost does not show up as a declined transaction; it shows up as a ceiling on aggregate quality that nobody can trace to its cause.
This is why we built Forge as a supply-side on-ramp: to lower the price of the scarce input by removing costs that have nothing to do with forecasting skill. A practitioner with a differentiated model should be able to become a worker on a topic without first becoming a protocol engineer. The narrower the non-forecasting gate, the more the contributor pool collapses toward people who happen to tolerate the plumbing, and the more correlated the surviving views tend to become. That is the opposite of what an aggregate benefits from. Every unnecessary gate we remove widens the distribution of approaches competing on a topic, and the width of that distribution is, in our analysis, the single variable most predictive of whether the aggregate will improve with the next worker who joins.
A network optimizing for quality should therefore measure itself on whether it is widening the set of approaches actually competing on its topics, not on aggregate compute or rows of data ingested. Those are the inputs the market already supplies cheaply. The one that determines the ceiling is the one that resists procurement, and it responds to accessibility rather than to spend. The implication for how we build is direct: every design decision that makes contributing easier for someone with a differentiated model and no interest in protocol plumbing is an investment in the input that actually moves the needle.
X The ceiling on a prediction network is not compute or data. Both are commoditizing. The scarce input is a model whose errors are not already covered by someone else on the topic. We built Forge to lower the cost of that input by removing everything unrelated to forecasting skill.
LinkedIn When we examine what limits the quality of a prediction network, the conversation usually centers on compute and data, because those are the inputs that are easy to count. Both have been commoditizing for years. Compute is rentable at the margin, and the historical series that short-horizon crypto forecasting depends on are largely shared across everyone bidding on the same topic.
What stays scarce is a genuinely different model: a contributor whose errors are not already reproduced by others on the topic. That input has no spot market. It tends to live with people outside crypto infrastructure entirely, carrying priors from signal processing, climate science, or an unusual feature set.
This is why we built Forge around accessibility. When the cost of contributing is fluency in wallets, gas, and epoch mechanics, the participants who clear the bar are selected for infrastructure tolerance, which is statistically independent of forecasting skill. The friction filters out the supply that would have improved the aggregate. A network serious about quality should measure how wide its set of competing approaches actually is, not how much compute it can marshal.
Get started with forge at https://forge.allora.network/
