Digital Signal Processing
Ehlers-class filters applied directly to canonical 1-minute axis. No resampling, no interpolation artifacts.
See the Engine.
Before You Commit.
Powered by PermuCheck™ and RunForward™ Engines
One million statistical tests are executed before a single backtest begins, completed in under sixty seconds. Exhaustive enumeration across the full parameter surface, with every configuration graded, filtered, and delivered.
Open Dojo →↓ Everything below runs live in Dojo
01 — THE PROBLEM
The Sequential Research Problem
Signal discovery, position sizing, and risk management are sequential dependencies. Optimizing all three simultaneously in a backtester is joint optimization over a massive parameter space with maximum degrees of freedom — the fastest path to overfitting.
1
Signal Discovery
Does this indicator crossover produce a statistically significant event? Across how many parameter combinations? Is it robust out-of-sample? Does it survive walk-forward, permutation null, and FDR correction?
WE DO THIS
→
2
Position Sizing
Given a validated statistical anomaly, how much capital per event? Kelly criterion, fractional Kelly, fixed-fraction — applied to signals that actually passed rigorous testing.
YOUR QUANTS DO THIS
→
3
Risk Management
Stop-loss, portfolio-level drawdown limits, correlation management across signals — informed by the statistical properties of vetted anomalies, not curve-fitted backtest outputs.
YOUR QUANTS DO THIS
The Coverage Problem
Traditional Approach
5–50
configs tested
0.00025% coverage
- Researcher picks favourite parameters
- Backtest with sizing + risk simultaneously
- Evaluate equity curve
- Eyeball an OOS split
- Ship it
VS
Student One
1,000,000+
configs enumerated
Complete response surface
- Every integer period 1–1,400
- 34 indicator families × crossover pairs
- Walk-forward, FDR, OOS & more gates
- Zero re-optimization on OOS
- Surviving anomalies delivered as feed
If you're a quant — you know you're skipping the hard part.
If you manage quants — you know how bad it looks. Your desk is doing backtesting in the name of research. Call it what it is: curve-fitting with extra steps.
You know what deserves ten thousand backtests? Permutations of position sizing. Variations on risk overlays. Execution timing. Those are legitimate optimisation targets — bounded parameter spaces with known distributions.
But signal discovery? That's not an optimisation problem. That's an enumeration problem. And you've been brute-forcing it with a backtester because nothing else existed. Now it does.
See how enumeration solves this in practice →Try in Dojo
02 — WHY THIS MATTERS
Win Rate Alone Is Meaningless
Different strategy types demand fundamentally different statistical profiles. The distribution of favorable vs. adverse excursion (MFE/MAE) determines whether a statistical anomaly is actionable — not just whether it has a positive win rate. You cannot assess this without exhaustive enumeration across the full parameter space.
Breakout / Volatility
Regime shifts via ATR/Bollinger expansionRequired Win Rate
28–40%
MFE / MAE Ratio
3–8×
Distribution Shape
Bimodal — clustered losses + rare large wins
This is the category most practitioners misread. A 30% win rate looks broken until you see the MFE distribution: the rare wins are multiples of the typical loss. If your out-of-sample MFE distribution narrows relative to in-sample, the regime shift you were capturing has decayed. The backtest number is irrelevant at that point. Without mapping the full MFE/MAE surface, you cannot distinguish a viable breakout signal from a curve-fitted artifact.
Without exhaustive enumeration, you cannot map the MFE/MAE distribution across parameter space. You are sizing positions on anomalies you have not statistically validated.
WHY BREAKOUT / VOLATILITY DESERVES ATTENTION
Most signal discovery workflows optimize for win rate. This creates a systematic blind spot: strategies where the edge comes from the magnitude of favorable excursion rather than the frequency of winning trades. Breakout and volatility expansion strategies fall into this category.
A configuration with a 31% win rate and an MFE/MAE ratio of 5.2 can outperform a configuration with a 68% win rate and an MFE/MAE ratio of 0.9. The former captures regime transitions. The latter captures noise. The only way to distinguish between them is to compute the full excursion distribution across the parameter lattice and observe which configurations maintain their MFE/MAE profile out of sample.
Enumeration surfaces these configurations because it does not pre-filter by win rate. Every configuration that satisfies the constraint set is returned, regardless of how its win rate appears in isolation. The bimodal distribution shape, the MFE clustering, and the MAE compression are all visible in the output. A human analyst or downstream model can then assess whether the trade-off profile fits the intended deployment.
The win rate ranges, MFE/MAE ratios, and distribution shapes shown above are approximate reference ranges derived from historical enumeration runs across multiple asset classes. They are illustrative of typical statistical profiles observed within each strategy category. They are not predictions, guarantees, or representations of future performance. Actual values depend on the specific asset, time period, parameter configuration, and market regime at the time of computation. All enumeration outputs delivered through the platform are computed on the specific data submitted by the user and may differ from the ranges shown here. These figures are provided for educational context only.
Run Win Rate, MAE, and MFE across all permutations →Try in Dojo
03 — ROBUSTNESS INFRASTRUCTURE
Advanced Statistical Robustness Gates
Every configuration that survives enumeration passes through a cascade of peer-reviewed statistical tests. No cherry-picking. No eyeballing. Each gate carries its academic citation and produces auditable metadata that your compliance team can reconstruct independently.
01
Win-Rate Gate
Cheapest filter. Percentage of events where close direction matched prediction.
02
Recurrence Gate
HHI-based concentration filter. Ensures signal is dispersed across the window, not clustered in a single burst.
03
Excursion Gate (MFE)
Mean Favorable Excursion threshold. Filters by magnitude of move in predicted direction.
04
Per-Regime MFE Gate
Regime-aware MFE with early termination. MFE must hold across volatility/session regimes.
05
Time-of-Day Buckets
Session-aligned temporal analysis. Crypto: 48 × 30-min UTC. Equity-RTH: 13 × 30-min ET.
06
Day-of-Week Mask
Include/exclude any subset of trading days. Per-day breakdown table in results.
07
Volume Confirmation
Events must fire on above-average volume. Formula: volume[t] > rolling_mean(volume, N=20) × k.
08
Volatility Regime
ATR(14) or Realized Volatility percentile band filter. Events must occur within specified regime.
09
Third-Indicator Regime Gate
The killer feature. Sweeps ~100 threshold points on a third indicator to find regime conditioning that improves signal.
10
Percentile Floor
Computes the q-th percentile (default 25th) of cost-adjusted per-event MFE. Must clear a basis-point floor.
11
Friction Sensitivity
Breakeven cost probe. Computes the round-trip cost (in bps) at which MFE crosses zero.
12
Exit Mix
Requires a minimum fraction of profitable exits from a specific reason (e.g. ProfitTarget).
13
Walk-Forward Survival
Pardo rolling validation: 180d train window, 30 rolls, 1d stride. Must survive ALL rolls.
14
Purged K-Fold CV
K-fold cross-validation with purge and embargo to prevent information leakage between train and test sets.
15
Probability of Backtest Overfitting (PBO)
Measures whether the best in-sample strategy systematically ranks below median out-of-sample.
16
MC Block-Bootstrap OOS
Monte Carlo block-bootstrap: randomly partitions time blocks into train/test across many iterations.
17
Hansen SPA
Superior Predictive Ability test. Controls data-snooping bias across the full survivor universe.
18
Permutation Null with BH-FDR
Shuffles returns K times (K=200 default, K=1000+ rigorous), applies Benjamini-Hochberg FDR at α=0.05.
19
Cluster Stability
DBSCAN on normalized parameter vectors. Annotation-only — never rejects qualifiers.
20
Auto OOS Split
Fixed last 20% of user window as out-of-sample. Zero re-optimization — in-sample thresholds applied as-is.
Apply any gate to your strategy. Validate statistical significance →Try in Dojo
04 — INDICATOR COVERAGE
41 Indicator Families. Every Integer Period 1–14,000.
Every crossover combination. Every threshold permutation. No cherry-picking. Applied to the canonical 1-minute axis — zero resampling artifacts, full statistical resolution preserved. Your traditional 5-minute or 1-hour bars? They're just subsets of what we test.
Oscillators & Mean-Reversion
Bounded indicators for distribution-extreme detection. Composite multi-factor confluence analysis.
Momentum & Adaptive Filters
Adaptive moving averages and trend-persistence quantification. Cross-asset divergence detection.
Volume & Volatility
Microstructure liquidity cycles, volatility regime detection, and flow analysis.
Canonical 1-Minute Sampling: Every indicator is computed on raw 1-minute OHLCV. Integer periods 1 through 14,000 are tested exhaustively. Higher timeframes are derived mathematically — no resampling, no information loss. This means a "daily RSI(14)" is actually tested alongside RSI(1) through RSI(14000) on the minute axis, revealing structure invisible to traditional charting.
Stop guessing RSI 14. Enumerate all 34 families against your asset →Try in Dojo
THE TECHNICAL ADVANTAGE
Why DSP-Based Sampling Matters
Traditional platforms resample OHLCV into fixed bars (1m, 5m, 15m, 1h) and lose everything between them. We don’t resample. Every integer period from 1 to 14,000 minutes is computed natively on a 1-minute canonical axis. Zero interpolation, zero aliasing, zero gaps.
✗ Traditional Resampling
Interpolation Artifacts · Data Loss · Discrete Gaps
- Missing: 2m, 3m, 4m, 6m, 7m, 8m… everything between fixed bars
- Resampling creates interpolation artifacts
- Phenomena between standard intervals = invisible
✓ DSP-Based Sampling
Zero Interpolation · Full Resolution · Every Integer Period
- Complete coverage: 1, 2, 3, 4, 5 … 720 minutes
- DSP primitives collapse to 1min canonical axis
- If a phenomenon exists, we find it at every integer frequency
CROSS-ASSET INTELLIGENCE
Gemina™
Divergence and convergence across every minute of every trading day, for any two assets, within or across asset classes.
Crypto × CryptoFX × FXEquities × CommoditiesAny × Any
Entire hedge funds are built around lead-lag and mean-reversion relationships between correlated instruments. We made this kind of information an API call away.
Run Gemina™ on any pair. Cross-asset convergence and divergence →Try in Dojo
DEEP DIVE
How It Works
A detailed technical reference for teams evaluating our statistical infrastructure. Everything below operates on the same engine that powers ESER™, SlipStream™, and the Machine API.
01
The Problem with Trusted Parameters
A researcher tests RSI(14) × RSI(70). Maybe RSI(7) × RSI(30). Perhaps five more combinations. That's 7 out of 196,000,000 possible configurations for a single indicator pair with periods 1–14,000. Coverage: 0.0000036%.
The parameter you trust came from a sample of effectively zero. The parameter space is not something you can manually explore. It's not something your quants are too lazy to do — it's computationally intractable without dedicated infrastructure.
02
PermuCheck — Single-Asset Indicator Sweep
Up to 1,000,000+ statistical tests per sweep in under 60 seconds. Tests every integer period from 1 to 14,000 across all crossover pairs for a given indicator family. Operates on native 1-minute canonical axis — zero resampling.
1M+tests/sweep
<60sexecution
1–14,000period range
1-mincanonical axis
03
Gemina — Lead-Lag Behaviour, Quantified
What time do they diverge? Who leads? What's the divergence magnitude? How long until they converge? For every asset pair — within or across asset class. Crypto, forex, equities, commodities.
Auto session-aligned. Just pick and drop two assets. Multiple lookback periods tested simultaneously. Discovers convergence relationships that traditional correlation analysis misses entirely.
04
RunForward — Survival Engine (Advanced Robustness Gates)
Every qualifying configuration passes through a cascade of peer-reviewed statistical tests. These are not backtests — they're multiple testing corrections that separate genuine phenomena from noise.
Walk-Forward SurvivalHansen SPARomano-Wolf Step-DownCluster Stability+ many more — you pick your stack
Citations: Hansen (2005), Romano & Wolf (2005), Benjamini & Hochberg (1995), Pardo (2008), López de Prado (2018), Politis & Romano (1994).
05
DSP-Based Sampling — Math That Shows Its Work
Traditional Resampling
- Aggregate 1-min to 5-min, 15-min, 1-hour bars
- Information loss at each aggregation step
- Interpolation artifacts
- Different results at different resolutions
DOJO DSP Approach
- Compute on native 1-minute axis always
- Ehlers Super Smoother, Roofing Filter, Hilbert
- Period parameter = mathematical filter, not bar aggregation
- Full statistical resolution preserved
06
Interactive Filter Demo
Sample output from a PermuCheck sweep. Each row = one parameter configuration that passed all gates. Filter below to explore the response surface.
| P1 | P2 | Smooth | N | Days | Occ Days | Win% | MAE | MFE | + |
|---|---|---|---|---|---|---|---|---|---|
| 14 | 70 | 3 | 847 | 252 | 189 | 61.7% | — | — | ⋯ |
| 21 | 80 | 5 | 612 | 252 | 156 | 65.0% | — | — | ⋯ |
| 7 | 30 | 2 | 1203 | 252 | 234 | 57.9% | — | — | ⋯ |
| 28 | 120 | 7 | 389 | 252 | 112 | 66.6% | — | — | ⋯ |
| 5 | 21 | 1 | 1544 | 252 | 248 | 55.0% | — | — | ⋯ |
| 50 | 200 | 10 | 201 | 252 | 78 | 70.1% | — | — | ⋯ |
| 9 | 45 | 3 | 956 | 252 | 201 | 60.0% | — | — | ⋯ |
| 35 | 140 | 5 | 445 | 252 | 134 | 64.0% | — | — | ⋯ |
| 12 | 55 | 4 | 778 | 252 | 178 | 62.0% | — | — | ⋯ |
| 3 | 14 | 1 | 1891 | 252 | 251 | 52.1% | — | — | ⋯ |
10 configurations shown. In production, sweeps return 10,000–100,000+ qualifying rows.
07
Data Pipeline Tools
LIVE
Fetchrr
Market data fetcher. Crypto, equities, FX. Multi-exchange support.
BETA
Loadrr
OHLCV parser. CSV, JSON, Parquet ingest. Schema validation.
PLANNED
Lablrr
Feature engineering. ML-ready Parquet datasets from enumeration output.
Seen enough? The engine is live and waiting →Open Dojo
05. ENTERPRISE INFRASTRUCTUREComplete Parameter Space Enumeration.
Complete Parameter Space Enumeration.
Delivered as Infrastructure.
Four integration paths. One statistical engine. Full explainability.
Triple-Redundant Compute
Three independent implementations. Three languages. Three machines. Identical seeds. Grade results 1/3, 2/3, or 3/3 based on agreement. The math is the same. The bugs are different.
Reconstructible by Design
Every statistic carries: thresholds, window sizes, indicator periods, smoothing levels, random seeds, timestamps, version hashes. Your auditor rebuilds from scratch without contacting us.
Full Parameter Lattice. Complete Statistical Surface.
Every valid configuration in your parameter space is computed and delivered. No sampling, no shortcuts, no hidden selections. The full finite-domain surface is yours to inspect, audit, and act on under your specified parameters.
Embed exhaustive statistical enumeration directly in your trading platform. Single HTTPS endpoint. White-label ready.
- → Single API Call. POST OHLCV, receive exhaustive results
- → Zero Data Retention. ephemeral compute, nothing persisted
- → White-Label Ready. style results in your own UI
- → Sub-60s Response. 1M+ configs enumerated per request
◆ Premium Institutional
Enterprise Statistical Exploration Report™
The Complete Response Surface. Delivered.
Exhaustive finite-domain pattern enumeration. Deterministic combinatorial sweep across all valid parameter configurations. Advanced robustness gates. walk-forward, permutation null, FDR, and more. Feed format of the same enumeration engine.
→
Enterprise Enquiries
For tailored engagement models, volume access, API integration, or machine API keys, contact our sales team.
Ready to see it in your environment?
Dojo lets you effortlessly try our products. Allow our sales team to demonstrate what the same technology can look like inside your platform.