Open Research / Observable Evidence / Proprietary Implementation
Do not ask for trust. Build for verification.
Don’t trust the backtest. Check what happened after it. Research should be open enough to be challenged. Evidence should be open enough to be examined.
The minimum checks we run on every backtest before a strategy is considered.
We search a wide parameter space and look for a plateau where results hold up, not a single best point.
We compute beta-adjusted alpha for every grid pattern and report the share of patterns with positive alpha, not just the champion.
Selection uses only data up to a cutoff; evaluation uses the unseen period after it.
We repeat select-then-evaluate as time moves forward, across several in-sample window lengths, so one lucky window cannot carry the result.
We estimate market beta and evaluate what remains after market exposure is removed; strategies can and do fail this step.
We also test with an expanding training window, so results do not depend on one rolling window length.
We break results down by market regime instead of relying on a single average.
We rerun the research as if we stood at an earlier date, using only information available then, and evaluate what followed.
Strategies that pass early tests are still rejected when later tests fail; rejected research is kept on record.
After every backtest check, the researcher still knows the past; only signals published in advance and observed afterward remove that.
Additional tests specific to DM-Signal, beyond the core checks above.
Measures the CAGR positive-rate across the entire parameter space. Broadly positive results across the parameter space reduce the likelihood that performance comes from a single isolated parameter spike, but do not by themselves rule out overfitting.
source: context/robustness-verification-catalog.md §0
Directly measures whether a champion selected in the past has eroded over time, using p_det_roll/p_erosion across 6/12/24-month windows.
source: context/robustness-verification-catalog.md §4.1
Verifies whether out-of-sample performance reproduces across champions selected at different points in time, using selection timelines and IS/OOS comparison.
source: context/robustness-verification-catalog.md §4.2
Confirms whether a champion is an isolated spike or a broad peak, by comparing the performance drop-off across neighboring parameters.
source: context/robustness-verification-catalog.md §4.3
Examines whether the apparent edge holds up against the benchmark during extreme periods such as sharp drawdowns or rate-hike shocks.
source: context/robustness-verification-catalog.md §4.4
Confirms whether a fixed champion maintains better-than-random performance within each walk-forward fold's population, via fold percentile.
source: context/robustness-verification-catalog.md §4.5
Measures how well the top set is preserved as the calculation window is shortened, using overlap rate / Jaccard similarity.
source: context/robustness-verification-catalog.md §4.6
Checks whether alpha remains positive across Bull/Bear/Sideways regimes, exposing apparent strength that depends on a single market condition.
source: context/robustness-verification-catalog.md §4.7
Locks selection rules and parameters at a fixed point in time, with no undocumented changes afterward. This freeze point becomes the boundary against the still-unverified future.
source: Rule Freeze
Observes and recalculates only real-time results after the Rule Freeze, kept distinct from the backtest. Forward metrics are computed using only rows from the forward period.
source: Forward observation
Everything before the line is still history. Selection rules are frozen (Rule Freeze), and only results observed after that freeze count as forward, real-time evidence. Forward metrics are recalculated using forward-period rows only.
Historical market data may be corrected later by data vendors; recalculated research series may therefore differ from the original historical signal record. Data corrections do not justify rewriting the old publication record.
Detailed, per-event revision log: in preparation.
Portfolio selection is deterministic: the same inputs always produce the same selection. Candidates whose scores fall within a small tolerance band are treated as tied, and ties are broken by a fixed, economically meaningful sequence, so tiny numerical noise cannot flip the ranking. The internal tie-break keys and weighting rules are not published.
Methodology version history is in preparation.
Research log (adopted / rejected / under test / deprecated) is in preparation.
Price data: Yahoo Finance daily prices (via yfinance), with in-house split and dividend adjustment checks. Daily closes are cross-checked against EODHD and Tiingo. Risk-free rate: 3-month U.S. Treasury bill (DTB3). Benchmark: SPY. System of record for signals and returns: PostgreSQL.
DM-Signal uses a separation between public research and member-only implementation similar to AQR's — not equivalent to AQR.
DM-Signal publishes systematic research and rules-based signals for informational purposes. Past performance, including forward-only results, does not guarantee future results. Nothing on this page is investment advice.