How ApolloAlpha works, in plain English
You do not need a maths background to follow this. It explains how a forecasting model turns data into a probability in general, then how ApolloAlpha does it each week, and why you can check the result rather than take our word for it.
1. How a forecasting model works, in general
Think of a weather forecast. Forecasters look at past conditions, notice what usually followed them, and say "70% chance of rain tomorrow." They are not predicting one certain outcome. They are saying that across many days that looked like this, it rained on about 7 in 10.
A statistical forecasting model does the same job with numbers:
Scroll down slowly. The picture below changes as you scroll, and you can touch parts of it to see what they mean.
A general teaching illustration of how machine-learning forecasting works. It is not a picture of ApolloAlpha's data, model or settings, which we keep private. All numbers are made up.
Data goes in
A financial model can learn from many kinds of data. These are examples of the kinds of data a model could use. Which data ApolloAlpha uses is private.
Prices. How a market has moved: the highs, lows and closes of past days and weeks.
- Data goes in. A long history of what happened.
- The model looks for the pattern. It learns how the data has related to the outcome in the past, and turns that relationship into a calculation.
- A probability comes out. For a new situation, the calculation gives a number between 0% and 100%.
- The result is tested. The only honest test is on cases the model did not see when it was built. A model that merely memorised the past looks brilliant on old data and fails on new data.
2. What ApolloAlpha forecasts
Each week, for each market it covers, ApolloAlpha answers two questions about the week ahead: will last week's low hold as support, and will last week's high hold as resistance? Each answer is a probability, for example "80% the low holds."
3. What we publish, and what we keep private
We publish how well the forecasts did, in full: every score, every chart, and the audit. We do not publish which data we use, how it is prepared, or the model that turns it into a probability. That work is our own intellectual property, and the illustration above is deliberately general so that it teaches the idea without describing ours.
4. How a forecast is made each week
After each trading week closes, the model is run on the latest information and produces a probability for each market. It is retrained and updated regularly with new data, and every forecast uses only information that was available before the week it is forecasting. Then the forecast is published, and later the outcome is added to the record.
5. What comes out
Two probabilities per market per week, the chance the low holds and the chance the high holds, shown in the app with the market's own record next to them. The model does not tell you to buy or sell. It states how likely each level is to hold.
6. How we know whether to trust it
A probability is only useful if it is honest: when it says 70%, about 70% should happen. So every forecast is scored on weeks the model had never seen, and the scores are published: accuracy, calibration, Brier score, AUC and the confusion matrix, all explained in plain language in our FAQ. We also test for look-ahead (using the future by mistake) and had the method independently re-implemented and audited. The results, and what they do not show, are on the track-record page.
7. What it is not
- Not a signal service and not trading advice. It publishes probabilities only.
- Not a crystal ball. A 90% forecast still breaks about 1 time in 10.
- Not live-trading proof. The published record is a walk-forward test on history. Live results will be reported the same way once the service opens.
Past results do not guarantee future results. This is statistical model output, not financial advice.