Author
The indicator sets the ceiling that no model can break through. A linear regression on a high-quality indicator beats a deep neural network on a low-quality one. Most R&D effort is spent on the model, where the marginal returns are smallest. The bigger gains live in the inputs.
The scientific method for trading is an eleven-stage protocol with pass/fail gates. Most candidate strategies die in the middle. The survivors are the only strategies worth running. Codify the protocol, run every candidate through every gate, accept the low survival rate.
Simplicity is not aesthetic preference, it is statistical advantage. Each parameter inflates standard errors, multiplies the search space, and gives noise a new lever to be mistaken for signal. Simple models generalize because complex ones cannot. The default complexity is lower than people think.
A market prediction commits before the outcome. A market explanation chooses after the outcome. The first is hard and economically useful. The second is cheap and psychologically comforting. Most commentary is explanation dressed in the grammar of prediction.
Most published TA is unfalsifiable. Every claim that is not structurally circular can be made falsifiable through a seven-step transformation: operationalize the trigger, bound the prediction, quantify success and failure, specify the benchmark, pre-commit everything. The rest is vibes.
A positive backtest return proves nothing about predictive power. The return decomposes into a sum of exposure contributions plus residual edge. Most retail rules are 95% exposure and 5% edge. Six diagnostic tests separate the two. Without them, bias travels as alpha.
A trading rule's return in isolation is meaningless. Information appears only against a benchmark. A long-biased rule in a rising market collects free drift. The bias-matched random rule strips it out. The choice of benchmark is the choice of conclusion.
The null hypothesis for any trading system is "this rule has no edge." The system has to falsify the null to be worth running. Most do not. The trader who skips the null test is shipping hope as evidence and treating luck as predictive power.
Every trading claim is induction: a pattern inferred from past data and projected forward. The conclusion is never certain. A 70% hit rate from 1000 signals carries sampling uncertainty plus the deeper uncertainty that the future may not be drawn from the same distribution as the past.
A backtest is one draw inside one history. The market plays out once and there is no second universe to compare it against. Statistics can address sampling variability within history. Statistics cannot address the fact that history itself is a sample of size one.
A backtest is one draw from a distribution of possible outcomes. With typical daily data and a few years of history, the 95% confidence interval on annualized return spans tens of percent. The headline number is honest. By itself, it is almost uninformative.
A backtest is an experiment producing data, not a screenshot proving a strategy. Every backtest return splits into predictive power plus drift-times-long-bias. Without detrending, a null model, and a p-value, an equity curve is decoration, not evidence.