Technical indicator constructed with the last X hours of data.
A lookback period of Y days.
Buy signals at level A in the indicator; sell signals at level B.
X and Y chosen to maximize profitability of the signals across a set of ETFs and A/B levels.
Trade the indicator for day Y+1 for the ETFs with the clearest signals, thus making the assumption that the next period will not differ significantly from the past Y periods.
Each day, reconstruct the indicator parameters over the moving lookback period and select a new set of ETFs to trade based on performance over the lookback period.
Repeat this process using daily periods for the indicator and weekly lookback periods. Using 5 minute periods for the indicator and hourly lookback periods. Etc.
If you do that, you find that Y is a small number, a smaller number than would allow for traditional statistical significance testing.
Which is what makes developing trading systems so difficult: before the market behaves in a pattern that can be detected conventionally, the pattern tends to change.
Which means that patterns have to be detected non-conventionally:
Predictability is a market variable that fades in and out over time.
The key is finding the time frame(s) where markets are displaying predictability/regularity.
Which would make a trader operate sometimes like a scalper, sometimes like a position trader, and sometimes like an investor.
Trading failure is a result of trying to make markets fit into a style of trading, rather than fitting the trading style to current market conditions.