ForeGlass Learn · Companion lesson
Equal point error is a starting point. Learn how uncertainty, comparable information and repeated outcomes help us judge a forecast more carefully.
The presenter is a computer-generated character with an AI voice. The numerical scoring examples are invented for illustration. They are not additional forecast results.
A useful range balances width and missed outcomes. The video explains weighted interval scores with a clearly labelled example.
Match the measure, period, outcome release and deadline. Keep same-inputs and own-inputs comparisons separate.
Many forecasts of one outcome still represent one observed case. Repeated, comparable outcomes are needed for broader skill claims.
What the first results establish
The first published comparison supports case-level error comparisons. It does not establish overall forecasting ability. The video explains how stronger evidence can be built over time without turning an illustrative score into an accuracy claim.
Several forecasts can have exactly the same error. That tells us their point predictions ended up equally far from the reported outcome. It does not tell us whether they used the same information, were made at the same time, or expressed the same uncertainty. ForeGlass keeps those differences visible.
A probabilistic forecast includes more than a single number. It can declare a median and a range with a stated probability. In this illustration, three forecasts have the same median but different eighty-percent ranges. The observation is three. A very wide range can cover the outcome without being very informative. A narrow range that misses can be overconfident.
These are the same invented forecasts as before. All predict a central value of two. The observed value is three. Each range gives a lower and upper bound intended to cover eighty percent of possible outcomes. A says one to three. B says one point nine to two point one, so it misses. C says zero to four. The score combines central error, range width and a penalty for a miss. Lower is better: A scores best here. Declare the ranges before the outcome. One case does not establish overall skill.
To compare fairly, match the measure, target period, official outcome release and forecast deadline. Then choose a track. In the same-inputs track, all methods receive the same data. In the own-inputs track, each method uses its permitted information. Keep these tracks separate. Forecast dates and available information stay visible. Being earlier does not earn an invented bonus.
Start with a simple baseline chosen before the results are known. For example, it could repeat the latest observed value. Add the scores over the same cases, using the same weights. Here the baseline totals ten and the model totals eight: twenty percent lower loss, not twenty percent accuracy. If the baseline's total loss is zero, that percentage is undefined. Show the underlying losses anyway.
Ten forecasts of one outcome still give us just one observed case. Combining forecasts into an ensemble does not create another outcome. To learn about forecasting skill, we need results for more distinct outcomes over time. Record which forecasts share data or build on each other. Even different outcomes can move together during the same recession. Show both the number of forecasts and the number of distinct outcomes.
Keep the data version, forecast, date, outcome and scoring rule together. The forecasting application calculates the score. Comuvia SDK helps retain the records and their links, so another person can inspect the calculation. A matching hash proves that the bytes match. It does not prove the claim is true, that the information was available earlier, or that the forecasts were independent.
Our first published comparison covers one shared outcome. It supports case-level error comparisons, with the original limitations preserved. Broader forecasting-skill claims need repeated, comparable outcomes and properly declared uncertainty. Explore the ForeGlass Comparison page to see both the results and what the evidence can support.
