Educational analysis only. Not financial advice. Signals are statistical patterns only.Full disclaimer
How Krentium's Signals Are Measured
Methodology disclosure for Krentium's quantitative signal engines.
Headline figures are in-sample backtest results.
Overview
Orientation to the two engines, the published hold horizons, and what "validated" means on this site.
Krentium publishes two engines: Wyckoff Daily and Wyckoff Weekly. Both run on the 649-stock survivorship-corrected universe at the 20d forward-return horizon. Per-engine details live in Per-Engine Validation; multi-horizon results are documented in Multi-Horizon Disclosure.
"Validated" here means each engine is backtested on a 649-stock survivorship-corrected universe and subjected to a battery of statistical tests — Carhart 4-factor alpha, block-bootstrap confidence intervals, and walk-forward temporal train/test splits — at 95% significance with Bonferroni correction. Headline performance figures are in-sample; walk-forward results use disjoint train/test time periods.
Validation Framework
The nine statistical tests used across engines — walk-forward, block bootstrap, Carhart 4-factor regression, Monte Carlo, transaction-cost sensitivity. 8 are published with full proof in the public validation repo; 1 is retained internally for engine confidentiality.
The validation battery spans five test families. Daily Wyckoff at the 20-day horizon carries the most extensive stack — nine tests applied. Other engines apply subsets appropriate to their claim type and data availability. Per-engine results are reported in Per-Engine Validation.
Walk-forward temporal validation
Train and test on disjoint time periods. The engine is built on data ending at one date; out-of-sample performance is measured on data after that date. Walk-forward win rate and signal counts per engine are reported in the Per-Engine Validation section.
Block bootstrap confidence intervals
Resample contiguous blocks of returns to construct confidence intervals around point estimates. Block resampling preserves the autocorrelation structure that simple resampling destroys. Reported as 95% CI bounds on win rate.
Carhart 4-factor regression
Per-signal alpha measured against a Carhart 4-factor model (market, size, value, momentum) using Fama-French total-return data. Alpha, t-statistic, p-value, and N reported per engine. Significance threshold: 95% with Bonferroni correction.
Monte Carlo simulation
Two distinct Monte Carlo tests: stock-selection (does the engine pick stocks better than chance?) and timing (does the engine time entries better than chance?). Both report percentile rank against a null distribution.
Transaction-cost sensitivity
Headline numbers are net of realistic transaction costs. The sensitivity test recomputes win rate across a cost grid to confirm no published metric depends on a cost assumption below realistic bounds.
The full per-test pass/fail record at the 20-day horizon is documented in the 9-test extended validation stack in Wyckoff Daily (8 published with full proof; 1 retained internally for engine confidentiality). Multiple-testing correction is applied across the battery: Bonferroni-corrected, significant at α = 0.05.
Per-Engine Validation
Engine-by-engine validation. Both engines run on the same 649-stock survivorship-corrected universe. Each engine's claim type is shown in italic under its header.
Wyckoff Daily
Forward-return validation · engine emits state, user picks hold
20d published hold · live in production
Wyckoff Daily is the daily-timeframe signal engine in the Wyckoff framework. Validation universe: 649-stock survivorship-corrected (see Universe Construction). Same-ticker signals deduplicated inside a 28-day window. All figures in this section are on the 20-trading-day forward-return horizon.
Signal economics
Carhart 4-factor regression
262 signals (of the 4,276 deduped total) are dropped from the regression: Fama-French factor values unavailable at signal date.
Walk-forward + bootstrap
Walk-forward (temporal out-of-sample, train 2006-2015 / test 2016-2025) win rate: 59.0%. Block-bootstrap 95% confidence interval on the 20d win rate (shown in the stat grid above) excludes 50% regardless of block size.
9 of 9 statistical tests pass at the 20-day horizon
Expand to see the list.
Show the nine tests
- Carhart 4-factor
- Block bootstrap CI
- Walk-forward temporal
- MC stock-selection
- MC timing
- Survivorship-bias
- Multiple testing (Bonferroni)
- Transaction costs sensitivity
- Effective N (Bartlett)
28-day deduplication window
A 28-day per-ticker deduplication window is applied across the Carhart regression, block bootstrap, and regime analyses. Bartlett autocorrelation analysis on per-ticker signal residuals returns the first insignificant lag at 28 days. Bartlett-adjusted effective sample size: 1,560 (22.16× inflation factor over the naive count).
Bull vs bear regime
Bull regime Carhart alpha statistically significant at 95% (N = 3,781). Bear regime estimate underpowered (N = 493); bull–bear alpha delta not significant at 95%. Earlier published values from a smaller historical universe showed a larger bear-regime alpha; current values include previously-delisted tickers.
| Regime | N | Carhart α | t-stat | p-value | Significant at 95% |
|---|---|---|---|---|---|
| Bull | 3,781 | +0.570% | 4.08 | 4.6e-5 | Yes |
| Bear | 493 | +0.670% | 1.42 | 0.154 | No |
Bull–bear alpha delta +0.10 pp (pooled t = 0.20, p = 0.838). Regime definition: bull if Fama-French cumulative market return is above its 50-day SMA, bear otherwise.
Crisis drawdowns
| Crisis | Window | Engine | SPY |
|---|---|---|---|
| GFC 2008 | 2007-10-10 → 2009-06-23 | 15.45% | 56.47% |
| COVID 2020 | 2020-02-19 → 2020-03-23 | 18.53% | 34.10% |
| Bear 2022 | 2022-01-12 → 2022-03-07 | 9.38% | 25.36% |
Peak-to-trough drawdowns from the standalone simulation. SPY values are price-only, without dividends.
Risk metrics
Standalone simulation — Wyckoff in isolation
Wyckoff Daily, standalone. $100 seed. Max drawdown and volatility deferred. Multi-horizon equity curves at 5d / 10d / 20d / 40d holds are visible on the Accumulation Daily tab of /performance.
Wyckoff Weekly
Forward-return validation · engine emits state, user picks hold
20d forward-return horizon · weekly-bar signal emission
Wyckoff Weekly runs the same framework on weekly bars. Validation universe: same 649-stock survivorship-corrected set applied across 2006-2026 with the 28-day per-ticker deduplication window. Per-signal statistical validation is complete; the standalone portfolio simulation (10-slot, 20d hold) is now populated and visible on the Accumulation Weekly tab of /performance. Crisis drawdowns and Weekly-specific risk metrics remain deferred.
Signal economics
Carhart 4-factor regression
Walk-forward validation
Walk-forward win rate: 61.15% (5-fold temporal OOS). Win rate above 50% under both temporal slicing and serial-correlation-aware bootstrap (CI above).
Bull vs bear regime breakdown
Bull regime: Carhart alpha +0.984% per signal (N = 2,485, t = 4.07, p = 4.8e-5). Bear regime: N = 16 signals across 19.5 years; sample size prohibits meaningful alpha estimation.
Weekly regime classification uses a different validation-time convention from the daily engine. The 16-signal bear cohort is too small to support either a positive or negative alpha estimate. 32 additional signals fell past the regime classification window and were excluded from regime classification.
Blocks deferred to a follow-up compute pass
The standalone equity curve at the 20-day hold (and 5d / 10d / 40d as exploratory disclosure) is now populated and visible on the Accumulation Weekly tab of /performance. The following Weekly-specific blocks remain deferred.
- Crisis drawdowns
- Per-Weekly-distribution crisis-window peak-to- trough metrics. The Weekly equity curve does include a 20d-hold drawdown chart on
/performance; per-crisis attribution (2008 GFC / 2020 COVID / 2022 Bear) on the Weekly distribution is the deferred element. - Weekly risk metrics
- Weekly-specific Half-Kelly and monthly VaR / CVaR. Daily-tab values are reasonable proxies given identical 20d-hold mechanics; per-Weekly-distribution refinement is deferred.
Multi-Horizon Disclosure
Forward-return validation at four holding horizons. The full grid, including horizons where the engine does not pass.
The Wyckoff daily engine was tested at four forward-return horizons: 5 trading days, 10 days, 20 days, and 40 days. Each row below is an independent Carhart 4-factor regression on the 649-stock survivorship-corrected universe, with a block-bootstrap 95% confidence interval on the win rate at that horizon. All four rows are shown, including horizons where the engine does not produce a statistically significant edge. Verdicts follow a four-tier convention (fail, marginal, pass, strong pass) defined in the footnote below the table.
| Horizon | N | Win rate | 95% CI | Carhart α | t | p | Verdict |
|---|---|---|---|---|---|---|---|
| 5d | 4,297 | 53.29% | [51.35%, 55.20%] | +0.075% | 1.12 | 0.261 | FAIL |
| 10d | 4,291 | 55.40% | [53.25%, 57.53%] | +0.192% | 1.95 | 0.051 | MARGINAL |
| 20d★published hold | 4,014 | 58.40% | [55.90%, 60.70%] | +0.509% | 3.70 | 0.0002 | PASS |
| 40d | 4,246 | 60.88% | [57.67%, 64.01%] | +0.961% | 4.74 | 2.2e-6 | STRONG PASS |
Verdict tiers: FAIL — no significant edge at the 5% level. MARGINAL — edge near the threshold of statistical significance. PASS — clear statistical significance at 95% confidence. STRONG PASS — edge at p-values well below 0.001. The 20-day row additionally passes a 9-test extended validation stack (see Wyckoff Daily). 8 are published with full proof; 1 is retained internally for engine confidentiality.
Why 20 days is the published hold
The published hold is 20 days. 40d shows a larger per-signal alpha (+0.961% vs +0.509%) but is not adopted because the full 9-test validation stack is complete only at 20d.
Tested horizon vs published hold
Tested horizon: forward-return measurement interval. Every row in the table above is a tested horizon.
Published hold: portfolio-simulation exit convention. At 40 days these diverge — 40d is tested but not adopted as the portfolio-sim hold per the reasoning above.
Multi-horizon analysis for weekly Wyckoff is deferred to the weekly compute pass (weekly horizon: 20d only).
Sector Breakdowns
Per-sector performance for Wyckoff Daily and Wyckoff Weekly.
Yahoo 11-sector taxonomy. Threshold for reliable inference: N ≥ 30 per sector. Sectors below threshold flagged "N too small" and rendered with "—" for win rate, mean return, and median return. Aggregate metrics (coverage, overall WR) include all signals in the rollup; only per-sector display is gated.
Wyckoff Daily
Coverage 100.00%. Deduped signals N = 3,928. Overall WR 58.40%.
Sector-breakdown source set (N = 3,928) is 348 signals smaller than the canonical Wyckoff Daily deduped count (N = 4,276). The gap reflects two-stage gating: 262 signals dropped for Fama-French factor unavailability (matching the Carhart regression N = 4,014), plus 86 additional signals dropped for SPY-calendar alignment in the sector breakdown only. Aggregate WR reproduces canonical 58.40%.
| Sector | N signals | Tickers | Win rate | Mean return | Median return |
|---|---|---|---|---|---|
| Technology | 885 | 101 | 60.23% | +2.32% | +2.06% |
| Healthcare | 663 | 58 | 56.26% | +1.11% | +1.01% |
| Industrials | 645 | 61 | 58.60% | +1.32% | +1.43% |
| Consumer Cyclical | 537 | 58 | 61.08% | +1.94% | +1.64% |
| Financial Services | 520 | 63 | 53.27% | +0.92% | +0.67% |
| Consumer Defensive | 256 | 33 | 58.59% | +1.27% | +0.80% |
| Basic Materials | 219 | 27 | 61.64% | +1.77% | +1.86% |
| Communication Services | 193 | 20 | 60.10% | +2.00% | +1.87% |
| EnergyN too small | 10 | 2 | — | — | — |
28-day per-ticker dedup applied before sector grouping.
Wyckoff Weekly
Coverage 96.88%. Deduped signals N = 2,533. Overall WR 61.15%.
| Sector | N signals | Tickers | Win rate | Mean return | Median return |
|---|---|---|---|---|---|
| Technology | 512 | 76 | 63.67% | +2.02% | +2.04% |
| Consumer Cyclical | 459 | 67 | 57.08% | +1.01% | +1.31% |
| Healthcare | 385 | 42 | 62.60% | +1.94% | +1.61% |
| Financial Services | 330 | 50 | 61.52% | +1.66% | +1.86% |
| Industrials | 269 | 47 | 62.08% | +1.63% | +1.70% |
| Communication Services | 169 | 32 | 56.80% | +1.04% | +1.22% |
| Consumer Defensive | 139 | 25 | 58.27% | +1.88% | +1.23% |
| Basic Materials | 127 | 20 | 63.78% | +4.74% | +1.37% |
| Unknown | 79 | 21 | 70.89% | +3.71% | +2.67% |
| UtilitiesN too small | 27 | 6 | — | — | — |
| EnergyN too small | 22 | 12 | — | — | — |
| Real EstateN too small | 15 | 4 | — | — | — |
Tickers whose sector could not be resolved are shown in the Unknown bucket.
The Unknown bucket comprises tickers whose sector could not be resolved via SEC EDGAR or yfinance lookups; the elevated WR likely reflects selection effects on the unresolvable subpopulation rather than a tradeable cross-sector signal.
Universe Construction
Stock universe construction: backtest and live scanner.
Backtest universe
649 US stocks: 237 current S&P 500 constituents and 412 previously-delisted or demoted tickers. Survivorship correction is applied by including the 412 delisted constituents in the validation signal set. Same-ticker signals are deduplicated inside a 28-day window before any aggregate statistic. Date range covered: 2006-2026.
Pre-backtest sector exclusions
Energy, Utilities, and REITs are excluded from the backtest universe. Driven by commodity prices, interest rates, and regulatory factors rather than equity-specific dynamics.
Live scanner universe
The live scanner runs nightly on 225 symbols: 215 stocks and 10 ETFs (SPY, QQQ, DIA, IWM, VOO, VTI, XLK, XLF, XLE, XLV).
The 10 ETFs are scanned for signals. None have surfaced at daily or weekly timeframes to date.
ETF buying signals surfaced at daily or weekly to date: 0.
Live vs backtest reconciliation
Backtest universe (N = 649) and live scanner stocks (N = 215) are distinct constructs. Overlap: 200 stocks. Live scanner excludes 37 backtest survivors selected by a per-engine performance threshold. Live scanner includes 15 stocks not in the backtest universe (14 from a wider candle dataset; 1 ticker rename, MMC->MRSH).
| Engine | Pruned N | Pruned share | WR if pruned removed | WR delta | α / mean-return delta |
|---|---|---|---|---|---|
| Wyckoff Daily | 303 | 5.1% | 58.03% | +0.57 pp | +0.066 pp (α) |
| Wyckoff Weekly | 270 | 9.3% | 61.05% | +0.31 pp | +0.067 pp (α) |
Removing the 37 pruned tickers from the backtest signal set shifts published win rate by sub-percentage-point amounts on each engine.
Limitations
Documented limitations of the validation: CBE pre-merger gap, weekly cap-tier coverage, SPY price-only.
- CBE pre-merger corporate-action gap. 12 pre-merger Wyckoff signals fired on CBE Nov 2012 before Eaton acquisition close.
- Weekly cap_tier coverage gap (structural). 62.6% unknown — not resolvable via further SEC fetches. Weekly cap_tier breakdowns not surfaced.
- SPY benchmark is price-only (no dividends). With dividends, SPY total return is higher than the price-only series shown. Carhart regression uses Fama-French total-return data; reported alpha is dividend-adjusted.
Regulatory Classification
Krentium publishes investment-strategy recommendations within the meaning of § 86 Abs. 1 WpHG (Anlagestrategieempfehlungen) and complies with MAR Article 20 + Delegated Regulation (EU) 2016/958. Krentium is not an investment firm (Wertpapierdienstleistungsunternehmen) and does not provide individualised investment advice (Anlageberatung) within the meaning of MiFID II Art. 4(1)(4). Public legal disclosures are maintained on the following pages: