Fear & Greed Parameter Sensitivity and Out-of-Sample Failure
Sentiment strategies can look attractive because of threshold selection, holding periods, and overlapping observations. This note applies a common benchmark, costs, executable sequences, sensitivity grids, and an out-of-sample split.
Define the question before inspecting returns
A fixed-horizon event study asks whether forward N-day returns differ on sentiment-qualified dates. It is not a portfolio because consecutive dates can all qualify, creating overlapping observations. The dual-threshold mode is closer to a trading rule: enter below the buy threshold and exit above the sell threshold after a minimum hold.
Conditioned DCA asks a different question: how units invested on fear dates perform by sample end. The three modes should not be ranked with one headline number; the tool discloses cash-flow and exit rules so event averages are not mistaken for portfolio returns.
Benchmarks, costs, and executable sequences
Bitcoin's long-run appreciation makes many entry rules appear successful. Fixed-horizon outcomes must therefore be compared with buying on any date for the same horizon. Excess—not merely positive—return addresses whether sentiment adds information.
Fees and slippage are deducted on entry and exit. Maximum drawdown does not compound all overlapping events; it uses an executable sequence in which the next trade can enter only after the previous exit. This reduces sample size but creates a coherent path for compounding and drawdown.
Sensitivity grids and out-of-sample tests
If threshold 17 and a 93-day horizon dominate while neighboring combinations fail, the result is more likely historical noise. A sensitivity grid should seek contiguous regions rather than the best cell, and the research decision must be fixed before inspecting the test segment.
Kieran Lab splits the first 70% and final 30% chronologically. This is not a full walk-forward design and cannot remove regime change, but it tests whether results depend on an early cycle; performance is not calculated for an undersized test sample.