ವಿಷಯಗಳು · 7 / 7
Test Threshold Sensitivity
ನವೀಕರಿಸಲಾಗಿದೆ 22 ಆಗಸ್ಟ್ 2026 · 3 ನಿಮಿಷ ಓದು · 14 ವಿಷಯಗಳು
ಸರಳ ಅರ್ಥ
Threshold Sensitivity tests whether an RSI-below rule chosen on training history remains stable on later held-out sessions. It sweeps a declared threshold grid with purged walk-forward folds, an embargo and a seeded permutation comparison. It does not choose or recommend a trading threshold.
Why it is useful #
The tool exposes how much a tuned in-sample threshold degrades out of sample, whether folds repeatedly select the same value and whether the observed mean is unusual against randomized draws from the same return pool.
Where it appears and the workflow #
Open Threshold Sensitivity, select the exact exchange listing, RSI period, threshold range/step, forward horizon, folds, embargo, permutations, seed and lookback. Review the declared rule, surface and fold observation counts first; then compare out-of-sample mean, hindsight mean, degradation, stability and permutation p-value.
Calculation or source #
For each walk-forward fold, training data selects the grid threshold with its own best mean forward return. Observations whose outcome window could leak into the test boundary are purged, and an embargo separates training from held-out sessions. The selected threshold is scored only on data that fold did not see. A fixed seed makes the permutation comparison reproducible.
Prerequisites, inputs, period and unit #
Prerequisites are an exact listing with enough persisted adjusted daily closes to compute Wilder RSI, the forward horizon and at least two valid folds. Thresholds use RSI points from 1 to 99; returns, degradation, stability and p-value are percentages; horizon, embargo and lookback count completed sessions.
Worked example #
Illustrative only: five folds select thresholds 30, 30, 27.5, 30 and 25. Out-of-sample mean is 0.4%, hindsight mean 1.1%, degradation 0.7 percentage points and permutation p-value 38%. That sample gives weak evidence of a stable tuned level and does not recommend RSI 30.
What high and low mean #
Higher selection stability means more folds chose the same parameter, while higher degradation means honest choices gave up more versus hindsight. A lower permutation p-value means fewer randomized draws matched the sample result; it is evidence against chance under this test, not proof of future effectiveness.
Positive, negative and zero #
Positive or negative mean return describes the held-out observations under the stated horizon. Zero can be a genuine mean or degradation. A missing mean, stability or p-value means too few valid observations or a degenerate comparison and must not be shown as zero.
Limitations and common mistakes #
Testing many grids, periods, horizons or symbols still creates multiple-testing risk beyond one report. Results are sensitive to regime, costs, overlapping outcomes and corporate-action history. Common mistakes include picking a grid after viewing results, reading best-possible hindsight as attainable, or treating a low p-value as predictive accuracy.
Market-specific differences #
The method uses completed exchange sessions and exact-listing adjusted history. Holiday calendars, history depth, liquidity and price limits differ by venue. The same numerical threshold can produce different observation counts across exchanges without implying an engine inconsistency.
Suggested next steps #
Freeze one question before running it. Record every input and the seed, inspect per-fold train/test/purged counts, compare out-of-sample with hindsight and then rerun one predeclared alternative. Keep both results and state why neither is a recommendation.
Educational use only #
This guide explains descriptive research evidence. It is not investment advice, a price prediction, a recommendation, a claim of predictive accuracy, or an instruction to buy, sell, rebalance or place an order.