How ApolloAlpha is tested, and what the numbers mean.
All figures on this page, including the headline numbers, calibration and accuracy by year, are based on our Apollo Axiom 2 model, except the side-by-side comparison with Apollo Orion 1ApolloAlpha publishes weekly probabilities that last week's high or low will hold or break. This page sets out the results, how they were produced, how they were independently checked, and, just as important, what they do not show. Figures are from our own audit, last extended on 29 October 2026, now covering 163 markets: the original 159 plus the Straits Times Index and the Singapore banks DBS, OCBC and UOB. Every outcome in all 163 was re-derived from raw prices and every published statistic was recomputed independently. The headline figures are from our latest model, the Apollo Axiom 2 model (Institutional plan); the Apollo Orion 1 model is compared with it below.
The headline numbers
One "call" is one side (support or resistance) of one market for one week. The 70%+ calls averaged 81.5% stated confidence and were right 82.6% of the time. For comparison, a guess that simply repeats the most common outcome so far scores about 55%.
Out-of-sample periods differ by market. Each market needs a stretch of history to train the model before it can be tested, so not every market is tested from 2016 onward. January 2016 applies to long-established markets with long price histories, such as the broad indices (Nasdaq, S&P 500). Markets with shorter histories are tested from the first week they could be, so their out-of-sample period starts later; each market's own start date is shown on its dashboard.
Are the stated probabilities honest?
A forecast is only useful if "70%" means about 70%. Comparing what the model said with what happened, across every call since 2016:
| MODEL SAID | AVERAGE STATED | ACTUALLY HAPPENED |
|---|---|---|
| 50-60% | 55.1% | 55.2% |
| 60-70% | 65.1% | 65.2% |
| 70-80% | 75.1% | 76.6% |
| 80-90% | 84.2% | 85.8% |
| 90%+ | 94.5% | 93.1% |
Stated and actual are close in every band: slightly under-confident between 70% and 90%, and slightly over-confident above 90%. We show it both ways on purpose.
Does it decay over time?
No. Year by year, accuracy stayed between 71.9% and 74.7% in every year except 2020, the weakest year at 67.7%. Over any 52-week stretch it ranged from 67.2% to 76.0%, and the latest 52 weeks are at 73.7%. Accuracy was 71.9% across 2016 to 2020 and 73.3% from 2021 onward, but that 1.4-point gap is within normal week-to-week noise, and 2020 alone pulls the earlier figure down, so we do not read it as an improvement. If the results had been overfitted to the past, later years would be expected to do worse, and they have not.
| YEAR | WEEKLY CALLS | RIGHT |
|---|---|---|
| 2016 | 14,256 | 73.2% |
| 2017 | 14,178 | 72.7% |
| 2018 | 14,350 | 73.1% |
| 2019 | 14,664 | 72.8% |
| 2020 | 14,805 | 67.7% |
| 2021 | 15,414 | 72.6% |
| 2022 | 15,514 | 74.7% |
| 2023 | 15,636 | 71.9% |
| 2024 | 15,890 | 72.9% |
| 2025 | 16,164 | 74.1% |
| 2026 (to date) | 13,614 | 73.6% |
Apollo Orion 1 model and Apollo Axiom 2 model, side by side
ApolloAlpha runs two models. The Apollo Axiom 2 model is our latest and the one behind the headline figures above. The Apollo Orion 1 model is our earlier model. Both are scored here on exactly the same markets and the same weeks, using the same prices, each market from the first week both models forecast it (1 January 2016 for most markets). Accuracy for the Apollo Orion 1 model uses its own decision threshold, so ranking quality (AUC) and probability error (Brier score) are the like-for-like measures.
Improvement = Apollo Axiom 2's relative improvement over Apollo Orion 1 (for the Brier score, the reduction in error), worked out from the rounded figures shown. Same 163 markets and the same weeks for both models, each market scored from its common out-of-sample start. AUC bars start at 0.5 (a coin flip); all other bars start at zero. Historical results, not a guarantee of future results.
How it is tested: no peeking at the future
- Walk-forward, out-of-sample. Each week's forecast uses only information available before that week. The headline record counts each market from its own out-of-sample start (January 2016 for long-established markets, later for markets with shorter histories).
- Look-ahead test. We cut the price history part-way, recomputed every earlier forecast using only the data before the cut, and compared. The probabilities were identical (largest difference 0.0) across 15 markets including SPY, QQQ, AAPL, XLE, oil, Bitcoin, HOOD, Broadcom, Eli Lilly and ExxonMobil.
- Independent re-implementation. The method was rebuilt from its written description, without using our production code, on freshly downloaded prices. Across the original 159 live markets and about 273,000 stored weekly forecasts, every probability matched (largest difference 0.0019, from solver tolerance) and every outcome matched (0 mismatches).
- Dashboard checks. Accuracy, Brier score, AUC, confusion matrices and charts shown in the app were recomputed from the stored forecasts and compared with what the app shows: 0 discrepancies.
What these numbers do not show
- This is a walk-forward test on history, not live trading. Live results start when the service opens, and we will report them the same way, wins and misses together.
- Accuracy is not profit. Predicting a level correctly does not make a trade profitable. Costs, entry timing, stops and position size all matter. ApolloAlpha publishes probabilities, not trade instructions.
- Some choices were made by the developer while looking at results. The model is small, with only three fitted numbers per market and side, but its design settings were chosen by us, not found by an outside party.
- Every market is scored from the same starting week for both models. That is 1 January 2016 for most markets, and the first week both models forecast it for the 31 markets whose history starts later, so the headline counts 164,485 calls rather than every call since 2016. The record covers 163 markets: the original 159 plus the Straits Times Index and the Singapore banks DBS, OCBC and UOB, added on 29 October 2026. For all 163 we re-derived every outcome from raw prices and recomputed every published statistic independently (0 disagreements, 0 mismatches). The look-ahead test and the full independent re-implementation above were run on the original 159 markets.
- The universe is today's list of markets, so names that no longer exist are not in it; the large-cap stocks added in October 2026 are today's largest S&P 500 companies. The effect on next-week level forecasts should be small, but it is not zero.
- Past results do not guarantee future results. This is statistical model output, not financial advice.
Check it yourself
Every forecast in the app shows its own record: accuracy, Brier score, AUC, the confusion matrix, a calibration curve and the recent calls, including the ones that were wrong. The free plan shows last week's forecast and how it played out for six markets, with no card needed. See how to read these statistics, or the plain-language guide How to predict whether support and resistance will hold.