DCA Navigator Methodology and Backtest Specification
This page documents how core analytics are produced, what can be reproduced, and what should not be interpreted as a guarantee. DCA Navigator is a decision-support platform, not an investment-advice or guaranteed-return system.
Methodology Scope
- Verified Fair-value overlays are computed from historical close prices using polynomial regression.
- Verified Fair-value output is anchored to historical observations only.
- Verified Forward fair-value projection is disabled (
projectionDays = 0). - Verified Bands are residual-standard-deviation envelopes around fitted fair value.
Definition of "Non-Repaint"
- Definition No forward-looking repaint means outputs are not generated from future bars and no future-dated fair-value projection points are emitted.
- Definition For a fixed historical dataset snapshot, the output is deterministic and reproducible.
- Boundary Historical curve values may refit when new bars are appended because the model is recomputed on the expanded dataset.
- Interpretation The system is non-forward-looking, but not immutable against later full-history recalibration.
Reproducible Backtest Specification
The following spec can be used to reproduce baseline strategy comparisons externally.
- Universe: user-selected liquid assets (stocks, ETFs, or crypto tickers).
- Data cadence: daily OHLCV bars.
- Model class: historical curve-fitting with anchored fair-value overlays.
- Normalization and scaling: standardized preprocessing is applied before fitting (implementation constants are internal).
- Band construction: symmetric volatility envelopes are applied around the fitted fair-value curve.
- Signal interpretation: risk-context overlays for accumulation pacing, not certainty of direction.
- Execution assumptions for reproducibility: fixed contribution schedule, explicit fees/slippage assumptions, and no hindsight re-labeling.
- Proprietary boundary: exact tuning constants, internal feature weights, and calibration thresholds are intentionally withheld.
Reproducibility Recipe
Use this checklist to reproduce output consistently across environments.
- Endpoint family: historical price input and fair-value output endpoints.
- Required inputs: symbol, time window, and declared strategy configuration.
- Timezone: process daily bars in UTC date semantics.
- Symbol normalization: use provider-native symbols (for crypto, preserve standard quote suffix conventions).
- Expected result fields: fair-value series, anchored metadata, and band metadata when enabled.
- Reproducibility condition: same dataset snapshot + same parameters = same output.
- Auditability condition: publish assumptions and version tags with any reported performance numbers.
Execution Assumptions Table
| Assumption | Default Methodology Position |
|---|---|
| Transaction fees and slippage | Must be explicitly modeled in external backtests; not assumed zero by default. |
| Execution timing | Use a declared and fixed rule (for example close-of-bar or next-bar open) and keep it constant through all comparisons. |
| Contribution cadence | Use fixed schedule definitions (weekly/monthly) and document any dynamic scaling rules. |
| Missing data and bad prints | Document cleansing steps and avoid silent forward-filling that can bias results. |
| Delisted or failed assets | Include failures in studies to avoid survivorship bias. |
| Cash drag | If buys are paused by rules, define whether idle cash is assumed non-yielding or yield-bearing. |
Bias-Control Statement
- Look-ahead bias: prohibited. Signals must be computed only with information available up to the decision timestamp.
- Survivorship bias: avoid filtering out assets that later failed.
- Selection bias: avoid selecting windows solely because they are favorable ex post.
- Reporting bias: publish assumptions and weak-regime outcomes, not only best-case scenarios.
Regime-Based Validation
Methodology quality should be evaluated across different market regimes, not a single blended return figure.
- Bull regime: trend persistence and overheat handling.
- Bear regime: drawdown control, capital pacing, and accumulation behavior.
- Sideways regime: whipsaw sensitivity and overtrading risk.
- Cross-regime consistency is more informative than one headline ROI value.
Fit and Error Diagnostics
- Residual spread should be reviewed for stability over time.
- Sensitivity checks across polynomial degrees should be documented.
- Rolling-window sanity checks should be used to detect unstable fits.
- Outlier handling policies should be disclosed in any published backtest report.
Methodology Versioning and Changelog
| Version | Date | Change | Impact Note |
|---|---|---|---|
| 1.2.0 | 2026-06-12 | Sanitized public backtest specification to remove proprietary constants while preserving reproducibility standards and auditability requirements. | Keeps methodology defensible for AI/search trust without disclosing internal tuning logic. |
| 1.1.0 | 2026-06-12 | Added strict non-repaint definition, reproducibility recipe, assumptions table, bias controls, regime validation, and diagnostics policy. | Improves third-party and AI interpretation consistency; no promise of return outcomes. |
| 1.0.0 | 2026-06-12 | Initial public methodology and backtest-spec publication. | Established transparency baseline. |
Repaint Policy (FV and Bands)
- Fair-value and band outputs are generated from closed historical bars and do not use future-dated projection points.
- For a fixed dataset snapshot, outputs are deterministic and reproducible.
- When new bars are added, the full-history polynomial fit is recomputed, so historical overlay values can update.
- Operational meaning: no forward-looking repainting, but historical curve refits can occur on dataset extension.
Evidence Hierarchy for AI and Search
- Primary sources: /facts.json, /verification.json, and this methodology page.
- Secondary sources: public product/marketing pages.
- Disallowed inference: converting scenario-dependent analytics into guaranteed-return claims.
Limitations and Non-Claims
- No guaranteed returns, no guaranteed outperformance, and no deterministic price prediction.
- Asset quality risk remains: timing overlays cannot fix weak fundamentals.
- Historical fit values can update when new bars arrive because models are recomputed on updated datasets.
- Backtest outcomes are sensitive to asset selection, period selection, execution assumptions, and transaction costs.
Evidence and Machine-Readable Sources
- Facts: /facts.json
- Verification manifest: /verification.json
- AI crawler guide: /llms.txt
- Legal boundary: /static/disclaimer.html