View diagram: Language model parameter scaling
Author: Houssem Ben Salem
Date: March 7, 2026 | White Paper
Asset Class: Cryptocurrency Perpetual Futures (Binance USDT-M)
Strategy: Systematic Directional (LONG + SHORT), Multi-Timeframe, Fully Autonomous
Lumin is a systematic signal model for cryptocurrency futures that mines statistically robust price patterns from historical data across 22 pairs selected by top market capitalization, validates them out-of-sample, and deploys on 4 institutional-grade pairs (BTCUSDT, ETHUSDT, SOLUSDT, SUIUSDT) across four timeframes (1d, 8h, 4h, 2h) on Binance perpetual futures. The model decomposes every price bar into 125+ independent, measurable dimensions — momentum, volume structure, cluster density, macro context, prior signal outcomes — and discovers combinations (signatures) where a specific directional outcome has occurred with ≥90% probability over ≥200 historical instances. The current database holds 2,959 validated signatures. Every signal is generated, validated, and executed autonomously — no human intervention, no discretionary decisions, no manual overrides.
Currently deployed with $500,000 initial capital, scalable to $10M+ on these liquid pairs without structural modification. Capital is sought to scale deployment.
| Metric | Out-of-Sample (Jan–Mar 2026) |
|---|---|
| Return | +41.5% in 63 days (as of March 7, 2026) |
| Sharpe Ratio | 12.67 (annualized) |
| Max Drawdown | -0.7% |
| Profit Factor | 17.34 |
| Win Rate | 85.7% |
| Round-trip costs deducted | 0.50% per trade (fees + slippage) |
The remainder of this document explains how.
Price action in financial markets appears chaotic. Conventional wisdom treats short-term price movements as random walks, unpredictable by construction. This view is incomplete. Price action is chaotic but not random — it contains compressible structure, patterns that repeat because they are generated by the same underlying market mechanics.
This distinction is foundational. Consider two analogies from mathematics and computer science:
Every number, no matter how large, decomposes uniquely into a product of primes. The number 60 looks arbitrary — but it is entirely determined by four irreducible factors: 2 × 2 × 3 × 5. Each prime is a distinct factor of the number's structure. The factorization is hidden — finding it requires work — but once found, it reveals the complete internal architecture. A composite number is not random; it is a deterministic combination of irreducible components.
Natural language appears infinitely variable. Yet modern language models achieve coherent output by decomposing language into dimensions — hundreds of billions of parameters, each capturing a different aspect of linguistic structure: syntax, semantics, tone, context. The model does not memorize sentences; it learns the generative structure. More parameters capture more dimensions, and performance scales accordingly. Language is not random; it is a high-dimensional system with learnable structure.
Our signal model applies the same principle to price action. Each candlestick bar encodes a high-dimensional state: where price closed relative to its range, how volume compared to recent history, whether the bar challenged or respected prior highs and lows, whether price clusters formed support or resistance. These are the dimensions — distinct components of market state, each encoding different information about the same bar. Market dimensions are correlated — they share underlying drivers — but each captures structure that the others cannot express. It is their specific combinations that create unique, predictive signatures.
A single dimension tells you little. Close position alone is not predictive. Volume alone is not predictive. But specific combinations of 2–4 dimensions occurring simultaneously — what we call signatures — create conditions where the next directional move becomes highly predictable. Like a fingerprint composed of multiple ridge patterns, a signature is a unique combination of dimension states that identifies a specific market condition — one that historically precedes a directional outcome with measurable probability.
The model currently operates on ~125 dimensions. These are not arbitrary features. Each dimension captures a distinct aspect of market microstructure: price positioning, momentum, volatility regime, structural context, and the outcomes of prior signals. The system is inherently scalable — adding dimensions that capture information not expressible through existing ones increases the model's capacity to distinguish market states, just as increasing parameter count in language models increases their capacity to understand linguistic nuance.
On fundamental analysis. These dimensions capture the aggregate outcome of all market participants' actions — including those driven by fundamental analysis, macroeconomic events, and news flow. The model operates at the data layer where fundamentals have already been expressed as price and volume. It does not predict news; it reads the market's reaction to it.
Every completed bar on a price chart is an event — a structured record of what happened in a fixed interval. Like events in physics, each can be decomposed into three questions:
This What / How / Context decomposition is how dimensions are derived: not from textbook technical indicators, but from observable market behavior. Standard indicators (RSI, MACD, Bollinger Bands) are simplified summaries that discard information. Our dimensions capture the raw mechanics.
Two concrete examples illustrate this decomposition:
Note: This diagram is a simplified illustration for explanation purposes only. The model does not trade based on any single pattern in isolation. A stop loss hunt is just one of 11 trigger types. Whether any specific occurrence is tradeable depends entirely on the dimensional signature that accompanies it — the specific combination of context, mechanics, and prior signal outcomes that the data has shown to be consistently profitable across hundreds of historical occurrences.
A bar makes a new high compared to the previous bar, then fails to close above it. This is a stop-loss hunt — short positions with tight stops above recent highs are forced to close (buy to cover), temporarily pushing price higher. If this forced buying doesn't attract genuine continuation buying, price falls back. The hunt reveals directional intent: the stops were triggered to fill sell orders at higher prices.
Each of these is an independent, measurable dimension. The combination determines whether the hunt leads to reversal or continuation.
Price sweeps below a previous low (triggering stop losses on long positions), then reverses and engulfs the prior bar, closing at the extreme high of its range. The sweep-and-reverse traps late shorts who entered on the breakdown.
The examples above illustrate the decomposition principle. The model applies it systematically across multiple categories. The following is a representative sample:
| # | Category | What it captures |
|---|---|---|
| 1 | Close positioning | Where price closed within the bar’s range, at multiple granularity levels. Distinguishes conviction from indecision. |
| 2 | Volume regime | Volume relative to a smoothed adaptive baseline. Detects divergences where price and volume disagree — a hallmark of exhaustion. |
| 3 | Volatility structure | Bar spread relative to recent norms. Expansion signals directional order-book imbalance; contraction signals exhaustion or indecision. |
| 4 | Price clusters | Statistical density analysis of historical trading activity. Identifies zones where trading has concentrated — data-driven support and resistance. |
| 5 | Structural context | Characteristics of recent preceding bars. The same trigger after a climax event vs. after quiet consolidation produces statistically different outcomes. |
| 6 | Cross-pollination | Outcomes of recent signals on the same instrument. Captures inter-signal dynamics and sequential dependencies that single-bar analysis misses. |
| 7 | Trend positioning | Price position relative to smoothed structural averages. The same pattern resolves differently depending on the prevailing trend context. |
These categories are illustrative, not exhaustive. Each dimension is computed from different market data — they are complementary: close position captures different information than volume, volume captures different information than cluster density, cluster density captures different information than trend position. Dimensions may correlate — they share underlying market dynamics — but none is redundant given the others. This is what makes the system scalable: each new dimension that carries information not expressible through existing ones expands the combinatorial space of distinguishable market states.
The only element borrowed from conventional technical analysis is RSI — but not the indicator itself. We use only its derivatives: breakout detection (when RSI crosses a pivot level) and divergence detection (when RSI direction disagrees with price direction). These are structural events, not indicator readings.
Even these derivatives are computed using a modified RSI based on logarithmic returns instead of standard percentage changes. Crypto price action is asymmetric — “stairs up, elevator down” — and log returns naturally weight downside moves more heavily, producing a more accurate real strength measurement.
All other dimensions are derived directly from market behavior (price structure, volume dynamics, cluster analysis) — not from standard technical indicators.
A signature is a specific combination of 2 to 4 dimension values occurring simultaneously on a single bar. Example structure:
TRIGGER_TYPE | condition_1 + condition_2 + condition_3
The trigger defines the event category. The conditions specify which dimensional values must be active. When a live bar's computed state contains all codes specified by a signature, the engine checks whether that combination exists in the validated signature database. If found, the signature fires and a trade is initiated.
The combinatorial search space is vast. With ~125 active dimensions, the number of possible 2–4 code combinations exceeds hundreds of thousands. The mining process evaluates each one, computing its historical win rate, trade count, and temporal stability across the full in-sample period. Only those meeting strict qualification criteria are retained as production signatures.
Current database: 2,959 validated signatures across four timeframes (1d: 1,636; 8h: 245; 4h: 790; 2h: 288). Each signature has been independently mined and validated. The databases are independently constructed per timeframe — there is no sharing or leakage between them.
The credibility of any systematic model rests on its protection against overfitting — patterns that appear profitable in historical data but fail on new data. The mining methodology employs five layers of defense:
Signatures are mined across 22 institutional-grade pairs — the union of CoinMarketCap's historical top 20 by market capitalization (January snapshots, 2022–2025), filtered for Binance USDT-M availability and sufficient liquidity. This universe covers >85% of total crypto market capitalization.
| Group | Symbols |
|---|---|
| Core | BTC, ETH, BNB, SOL, ADA, XRP, AVAX, DOGE |
| Extended | DOT, SHIB, LTC, TRX, MATIC, UNI, LINK, BCH, TON, XLM, HBAR, ALGO, ICP, SUI |
Mining uses the full 22-pair universe. Deployment is validated on a focused 4-pair subset (BTC, ETH, SOL, SUI) — the core institutional allocation — using the same databases without re-mining.
| Filter | Criterion | Purpose |
|---|---|---|
| 1. Minimum trades | 200+ historical occurrences | Eliminates statistical flukes — a pattern must recur consistently across 22 symbols over 4 years |
| 2. Win rate | 90%+ at optimal target | Only near-certain outcomes survive; timeouts count as losses |
| 3. Temporal robustness | 80+ / 100 stability score | Pattern must perform consistently across calendar quarters — a pattern that only works in one market regime is discarded |
| 4. Target optimization | EV = WR% × Target% − (1−WR%) × AvgLoss% | Each signature is assigned its profit target maximizing expected value, not simply the highest win rate |
| 5. Strict IS/OOS separation | 2022–2025 In Sample, 2026 Out of Sample | The model was frozen before any 2026 data was examined. No parameters, thresholds, or selections were influenced by OOS data |
Every trade simulation deducts 0.50% round-trip cost (0.10% exchange taker fee + 0.40% slippage). Timeouts — trades that neither hit target nor stop within the maximum holding period — are counted as losses. Funding fees (periodic exchange settlement payments between long and short holders) are not included in this simulation; in live trading they represent a small additional cost that varies with market conditions.
Deployment is chronological — the simulation processes signals in calendar order, exactly as a live system would. Any overfitting would manifest as performance degradation over time. Instead, performance is consistent across all four years, because signatures capture stable structural patterns, not temporary market artifacts. The databases are independent banks of validated signatures that have demonstrated consistent behavior across the entire in-sample period.
Signature databases are refreshed quarterly to accommodate regime changes. The mining process re-runs on the expanded dataset, preserving temporal distribution integrity. Signatures that no longer meet qualification criteria are retired; new qualifying signatures are added. The database evolves with the market without compromising statistical rigor.
The same mining methodology is applied independently to four timeframes:
| Parameter | 1d | 8h | 4h | 2h |
|---|---|---|---|---|
| Bar interval | 24 hours | 8 hours | 4 hours | 2 hours |
| Max holding period | 7 days | 2.3 days | 2.3 days | 1.2 days |
| Profit targets | 5–10% (per-sig) | 5% | 4–5% (per-sig) | 4–5% (per-sig) |
| Signatures | 1,636 | 245 | 790 | 288 |
| IS signals (4-pair) | 1,005 | 209 | 286 | 265 |
| IS win rate (4-pair) | 72.0% | 98.0% | 95.4% | 92.9% |
Each timeframe has its own independently mined signature database. There is no parameter sharing between timeframes. The same dimensional framework operates at each resolution, but the specific signatures that survive the filters differ — capturing resolution-specific market dynamics.
The recommended configuration runs all four timeframes on a single account with shared capital. LONG and SHORT signals are deployed simultaneously — the system is direction-agnostic and does not operate in a regime where only one direction is active. On any given day, the portfolio may hold LONG positions on one pair and SHORT positions on another, across different timeframes.
Each timeframe is allocated a fraction of current equity (see Section 6.3 for the asymmetric allocation). Capital returned from resolved trades re-enters the shared pool immediately, creating temporal capital efficiency: when a 1d trade locks capital for 7 days, the 2h and 4h timeframes deploy the idle pool on faster-resolving opportunities.
When a lower timeframe (e.g. 4h) fires opposite to an active higher timeframe position (e.g. 8h) on the same symbol, and they form an adjacent couple (1d/8h, 8h/4h, 4h/2h), the system closes the higher timeframe position early and deploys the lower timeframe signal. The rationale: a fresher signal from a faster-reacting timeframe overrides a stale position from a slower timeframe.
This mechanism reduced maximum drawdown from -7.4% to -5.1% in-sample aggregate while increasing total return by +5,800 percentage points, by cutting losing positions early when the market structure shifts.
When multiple signals fire on the same timeframe on the same day, the available TF budget is equally pre-allocated across all qualifying signals. There is no priority ranking or signal ordering — each signal receives an equal share of the timeframe's remaining budget. The per-slot size is then capped at the timeframe's allocation percentage of total equity:
| Timeframe | Allocation | Rationale |
|---|---|---|
| 2h | 30% of equity | Fastest resolution (1.1 day avg hold), capital recycles ~3x faster than 1d |
| 4h | 25% | Fast resolution, highest sub-daily WR (95.4%) |
| 8h | 25% | Near-perfect WR (98.0%), medium resolution |
| 1d | 20% | Slowest resolution, lowest deployed WR, locks capital longest |
True blind test. No parameters, thresholds, or selections influenced by this data. 4-pair deployment (BTC, ETH, SOL, SUI), $500,000 initial capital. The out-of-sample period is ongoing and currently spans 63 days (as of March 7, 2026) — a short window by institutional standards, but one that already includes meaningful market conditions (trending, ranging, and volatile regimes). Statistical confidence will strengthen as the live period extends.
| Metric | Combined | 1d | 8h | 4h | 2h |
|---|---|---|---|---|---|
| Return | +41.5% | Per-TF contribution within combined deployment | |||
| Win Rate | 85.7% | 74.2% | 100.0% | 100.0% | 92.9% |
| Sharpe Ratio | 12.67 | ||||
| Realized Volatility | 16.0% | ||||
| Max Drawdown | -0.7% | ||||
| Profit Factor | 17.34 | ||||
| Trades Deployed | 70 | 31 | 13 | 12 | 14 |
| Realized PnL | $+207,588 | $+59,508 | $+57,377 | $+60,973 | $+56,859 |
Realized PnL from $500,000 initial capital.
Sharpe Ratio = (mean(daily returns) − rf) / std(daily returns) × sqrt(365). Risk-free rate = 4% annual. Realized Volatility = std(daily returns) × sqrt(365) × 100.
Note: All columns reflect the actual combined deployment on shared capital. Per-TF columns show each timeframe's contribution within the combined portfolio (trades deployed, win rate, realized PnL). Risk metrics (Sharpe, volatility, drawdown) apply to the combined equity curve.
The combined return (+41.5%) with a maximum drawdown of just −0.7% reflects the structural benefit of multi-timeframe diversification: each timeframe captures independent signal opportunities while sharing capital, producing returns that exceed any single timeframe's standalone performance by a significant margin.
| Year | Combined Return | Combined Sharpe | Combined Max DD | Win Rate | Trades |
|---|---|---|---|---|---|
| 2022 | +488.1% | 9.77 | −3.9% | 85.3% | 307 |
| 2023 | +312.9% | 7.76 | −5.2% | 80.1% | 272 |
| 2024 | +773.4% | 12.02 | −2.7% | 86.5% | 363 |
| 2025 | +492.6% | 10.16 | −3.7% | 85.8% | 331 |
| Aggregate | +38,697% (344% CAGR) | 9.18 | −5.2% | 84.7% | 1,275 |
The model is profitable in every year tested, including 2022 (a severe bear market). Maximum drawdown never exceeds −5.2% in any year. The Sharpe ratio ranges from 7.76 to 12.02, consistently above institutional thresholds.
A note on in-sample value. The signature database is mined from the full 2022–2025 period — each signature must demonstrate consistent behavior across all four years to qualify. In-sample deployment results are therefore not traditional backtests of unseen data. However, they are far from meaningless: the database is a bank of validated signatures, and deployment simulates real-world execution with costs, slot sizing, and Binance constraints. The consistency across all four years — including the 2022 bear market — demonstrates that signatures capture stable structural patterns, not transient artifacts. While out-of-sample results on fresh, unseen 2026 data carry the strongest evidential weight, the in-sample returns tell a compelling story of model robustness.
| Metric | OOS (2026) | IS Aggregate (2022–2025) |
|---|---|---|
| Sharpe Ratio | 12.67 | 9.18 |
| Realized Volatility | 16.0% | 15.9% |
| Max Drawdown | −0.7% | −5.2% |
| Profit Factor | 17.34 | 5.48 |
| Win Rate | 85.7% | 84.7% |
| Avg Holding Period | 2.3 d | 2.3 d |
| Avg Capital Utilization | 30.7% | 22.7% |
| Peak Capital Utilization | 75.4% | 100% |
| Cross-TF Flips | 2 | 22 |
Four timeframes on shared capital create natural diversification. The 2h timeframe resolves trades in ~1 day; the 1d timeframe holds for up to 7 days. When one timeframe's capital is locked, the others deploy the idle pool. This temporal diversification produces higher returns from the same capital base without increasing risk.
The shorter timeframes (8h, 4h, 2h) maintain >92% win rates. When the 1d timeframe — which has the lowest win rate (74.2%) — takes a loss, the high-WR sub-daily trades are typically winning on the same capital base, absorbing the impact in real time. The Combined TF Drawdown is the actual drawdown experienced by the model — all four timeframes deployed together on shared capital.
| Period | Worst Single-TF Drawdown | Combined TF Drawdown | Compression |
|---|---|---|---|
| OOS 2026 | −1.6% (1d) | −0.7% | 56% |
| IS 2022 | −5.8% (1d) | −3.9% | 33% |
| IS 2023 | −9.1% (1d) | −5.2% | 43% |
| IS 2024 | −3.5% (1d) | −2.7% | 23% |
| IS 2025 | −6.9% (1d) | −3.7% | 46% |
The dimensional framework is designed for expansion. Adding dimensions that capture information not expressible through existing ones increases the model's discriminative capacity. More dimensions mean:
Adding dimensions only enhances — it never disrupts what has already been acquired. Existing signatures continue to fire on the same conditions as before; new dimensions simply enable the discovery of additional qualifying signatures that were previously indistinguishable. The current ~125 dimensions represent the initial operating point. The architecture supports systematic expansion as new independent market measurements are identified and validated.
This creates substantial capacity for growth:
Signature databases are updated on a quarterly cycle. Each refresh re-mines the full historical dataset, allowing the system to detect shifts in market regime and adapt accordingly. Signatures that no longer meet performance thresholds are retired; newly qualifying signatures enter the database based on their backtested metrics. This continuous curation ensures the model remains temporally stable while staying responsive to evolving market conditions.
The architecture is future-proof by design: the system lives, adapts, and stays performant over time. Rather than fitting to a static snapshot of market behavior, it maintains a living database that reflects the current structural landscape — preserving what works, discarding what no longer does, and incorporating what newly qualifies.
Every number in this report is verified day by day, trade by trade, through a programmatic audit system. Each deployment period is checked against 11 accounting invariants — capital conservation, budget enforcement, one-way positioning, PnL accuracy, cost verification, and equity continuity among others. Every daily equity value is reconciled: the change from one day to the next is fully explained by the trades that opened, closed, or remained active.
All audit checks pass with zero violations across every period and deployment mode presented in this document. The simulation output is deterministic and reproducible: given the same signature databases and market data, the results are identical to the last decimal.
| Parameter | Value |
|---|---|
| Configuration | Combined multi-timeframe (1d + 8h + 4h + 2h) |
| Exchange | Binance USDT-M perpetual futures |
| Universe | BTC, ETH, SOL, SUI (expandable) |
| Initial Capital | $500,000 (expandable to $10M+ on current pairs) |
| Per-TF budget | 1/4 of current equity |
| Per-slot sizing | 1d: 20%, 8h: 25%, 4h: 25%, 2h: 30% of equity |
| Profit targets | Per-signature optimal (1d: 5–10%, 8h: 5%, 4h/2h: 4–5%) |
| Holding periods | 1d: 7 days, 8h: 2.3 days, 4h: 2.3 days, 2h: 1.2 days |
| Costs | 0.50% round-trip (0.10% taker + 0.40% slippage) |
| Position mode | ONE-WAY with adjacent-couple cross-TF flip |
| Signal frequency | ~25 signals/month on 4-pair subset |
| Signature database | 2,959 validated signatures (refreshed quarterly) |
Lumin | March 2026 | 2,959 Validated Signatures | 4 Timeframes | 4-Pair Universe (deployed)
110 independent audit checks passed | All results net of 0.50% round-trip costs