The previous lesson said which indicators to keep. This one shows the working, including the results we did not want.
Everything below is from 67,098 point-in-time observations, 2016-01 to 2026-06, split walk-forward: designed on 2016-2022, then run untouched on 2023-2026. The outcome measured is touched +20% within 30 sessions.
The full scoreboard
| Indicator reading | TRAIN | TEST | Verdict |
|---|---|---|---|
| Close above 52-week high | 1.91× | 1.29× | survives |
| Fallen >50% from 52-week high | 1.74× | 1.26× | survives |
| Breakout on 2×+ volume | 1.63× | 1.26× | survives |
| RSI above 80 | 1.60× | 1.35× | survives |
| Breakout + above 200-day average | 1.55× | 1.25× | survives |
| Riding above the upper Bollinger band | 1.36× | 1.21× | survives |
| Bullish divergence | 1.18× | 1.66× | unstable |
| RSI below 30 | 1.36× | 1.09× | decayed |
| Bearish divergence | 1.47× | 1.06× | decayed |
| Above the 200-day average, alone | 1.10× | 0.98× | nothing |
| Above the 50-day average, alone | 1.08× | 1.00× | nothing |
| Bollinger squeeze | 0.88× | 0.90× | worse than random |
| At the 20-day low ("support") | 0.86× | 0.87× | worse than random |
How to read a table like this properly
The important column is TEST. The years the rules had never seen. Look at what happens to the gap between the columns:
- Rules that survive keep the same sign and roughly the same shape.
- Rules that decay (RSI<30, bearish divergence) looked strong in the design years and faded. That is the signature of a rule that was fitted to an era.
- Rules that show nothing flip sign, 1.10× becomes 0.98×. A flip is not a weak signal, it is an absent one.
This is why any performance claim without a held-out period should be ignored. Every rule in the "decayed" block would look excellent if we had only shown you TRAIN, and every one of them would have cost you money.
The finding underneath all of them
Every surviving row describes a stock already in motion. Every failed row describes a quiet or mildly falling one. Six indicators, one conclusion, which is itself a good sign, because independent tools agreeing is more convincing than any single number.
And the honest limits
Observations overlap in time, so the effective independent sample is smaller than 67,098. Delisted companies are absent from our price history, which flatters every figure. Entries are modelled at the close. And two periods agreeing does not guarantee a third will.
Do this now, practical activity
- Write down the three indicators you currently rely on most.
- Find each one in the table above.
- For any that lands in "nothing" or "worse than random", write one sentence on what you will do differently.
If none of your three appear: that is genuinely interesting and worth telling us. It means you are using something we have not tested, and untested is not the same as wrong.
Measured: across 67,098 observations walk-forward, 6 classic readings survived both periods, 2 decayed, 2 measured as nothing, and 2 measured below the base rate in both periods.