How to predict whether support and resistance will hold
Most traders draw a support line under last week's low and a resistance line over last week's high, then treat them as walls. Sometimes price bounces off them. Sometimes it blows straight through. The useful question is not "is this a support level?" but "how likely is it to hold?"
1. Support and resistance, in one paragraph
Support is a price level where a market has recently stopped falling. Resistance is a level where it has recently stopped rising. A common, simple definition is last week's low as support and last week's high as resistance. A level holds if the following week does not trade through it (a low that holds is not undercut; a high that holds is not exceeded), and it breaks (a breach) if the week does trade through it.
2. Think in probabilities, not lines
Every level has some chance of holding. A forecast that says "80% this low holds" is claiming that, across many similar situations, about 8 in 10 hold and 2 in 10 break. That is more honest than a hand-drawn line, and it lets you compare levels: a 90% level and a 55% level should not be treated the same.
3. How to check whether a forecast is reliable
Do not trust a headline accuracy number by itself. Ask for these:
- Out-of-sample testing. Was it tested on weeks the model had never seen when it made each forecast?
- Calibration. When it said 70%, did it happen about 70% of the time?
- Brier score. It punishes confident wrong calls. Always saying 50% scores 0.25; lower is better.
- AUC. It shows whether the model ranks good weeks above bad ones. 0.5 is a coin flip; 1.0 is perfect.
- A baseline. Beating "always guess the most common outcome" is the minimum bar.
Each of these is explained, with diagrams, in the statistics section of our FAQ.
4. Mistakes that make results look better than they are
- Look-ahead bias. Letting a forecast use information from after the week it is predicting. A simple test: cut the data at a date, recompute earlier forecasts, and check they do not change.
- Overfitting. Tuning a model until it fits the past perfectly. If accuracy fades in later years, this is the usual cause.
- Cherry-picked markets or periods. Show results for every market and for the whole period, including the misses.
- Confusing accuracy with profit. Calling a level correctly is not the same as making money from a trade. Costs, timing and position size all matter.
5. Where ApolloAlpha fits
ApolloAlpha (also written Apollo Alpha) publishes this kind of probability every week for 160+ markets: the chance that last week's high or low holds or breaks, with the record behind it. Across 164,485 weekly calls tested out-of-sample from January 2016, 72.7% were right, and 82.6% were right when the model was at least 70% confident. These are results of a historical test, not live trading, and past results do not guarantee future results. See how the record was tested and audited, and what it does not show.
This is statistical model output, not financial advice.