Frequently
asked.

We are a well-funded cybersecurity company operating in stealth. Our systems excel at pattern detection, analysing terabytes of data across vast numbers of data points.

This forecasting website is our pet project: a place where we play with forecasting the future, partly using a subset of our research and prediction systems.

We mostly choose questions that we find interesting. Some come from readers. Others come to us in the shower, or while browsing the web when we should be working.

The most important question is always: are we interested in knowing the answer? If the answer is no, we do not bother.

1. Finding the right idea. Start with a question worth answering. Readers can suggest one here.
2. Researching the data. Using our automated systems, we search for relevant data. The information is then embedded in a large data warehouse and organised into data points across many facets and dimensions.
3. Running the prediction model. The data points are fed into a model whose one and only area of expertise is answering a Boolean question—true or false—and assigning a probability to the answer. The model also produces a matrix identifying the relevant data points and showing how each one affected the decision.
4. Generating the draft. The matrix is fed into a text-generation pipeline, which produces the first draft of the article.
5. Adding the humans. Every article is read by someone on our team before publication—usually someone who genuinely wants to know the answer. This is also the stage at which biased human opinions are injected into the text.
6. If you are looking for 100% unbiased forecasts, well… this is not the right place. We believe that machine brains must always be governed by human hearts.

A ParallaxSee forecast is a fusion of human and machine intelligence.

Humans select the interesting questions. Machines perform sophisticated data aggregation, fusion and inference. Then we add the secret sauce: human perspective, curiosity, love and interest.

The percentage is the estimated probability that the forecast will resolve as true.

Percentages may change over time. When new data points are added to the knowledge base, the model may produce a different estimate. If the anticipated change is large enough, the published probability may be updated.

The article itself may occasionally lag behind the latest probability because we are currently more focused on publishing new content. We expect to refresh older articles more frequently in the future.

The two articles make opposite predictions using the same deadline and resolution boundary.

Our first reader-submitted forecast argued that Andy Burnham would not remain Britain's prime minister until the next general election. We wrote the forecast because we found the question interesting.

Then we decided to write the opposite forecast because… well… we really want him to remain prime minister.

Because the two outcomes are opposites, their probabilities are 26% and 74%.

A Reader's Forecast begins with a forecast submitted by a reader.

We choose the submissions that interest us, preserve the reader's central arguments and then enhance them using the research and writing pipeline described above.

The forecast remains attributed to the reader. Enhancing it does not silently turn it into the House position.

The target year is the forecast's deadline. The resolution language in the article explains exactly what must happen by that point.

A technology being demonstrated is not necessarily the same as its commercial adoption. A book announcement is not the same as publication. Winning a nomination is not the same as becoming president.

The resolution test controls the result.

The available evidence is reviewed against the forecast's original resolution conditions.

The forecast can then be marked correct, incorrect or unresolved if reliable evidence is still unavailable.

God, no!

Our state-of-the-art cybersecurity systems can predict bot attacks, stop them and identify bot traffic even when it is encrypted. When enough data is available, we can determine with 99.5% accuracy whether a website user is a bot.

We can do this for a simple reason: within the networking domain, we can access nearly all the relevant data required to verify or falsify the claim that a user is a bot.

The human world is infinitely richer than the networking domain. The “real-world data” available to our research systems is only a tiny subset of everything that could affect an outcome.

Truthfully, predicting the future is like describing the Mona Lisa while looking at it from a great distance through a narrow cylinder. It is challenging and fun—but not particularly accurate.

No.

Ten articles repeating the same company press release still represent only one underlying claim.

In simplified terms, the calculation is closer to:

Duplicate or dependent data points are identified and removed from the calculation wherever possible.

In other words, more good, reliable and unique data is helpful. Nothing less and nothing more.

Yes. Disagreement is part of the public record.

We call a well-founded and clearly articulated disagreement an intervention, and we love receiving them.

Readers can challenge an assumption, introduce new evidence or construct a counterforecast with a different probability. The objective is not to eliminate disagreement, but to make it specific enough to examine.

No.

ParallaxSee forecasts are editorial research about possible futures. They should never replace professional advice or individual decision-making.

A forecast can remain useful if its claim, evidence and probability were stated clearly. A failed prediction reveals which assumptions broke and helps improve the next forecast.

The main purpose of every forecast is to be interesting and to provoke an internal or public debate. Of course, we would love to be right all the time—but the future business is a risky business.