Abstract
view on ssrnQuantitative risk modelling and predictive alpha generation in high-volatility financial markets, exemplified by Bitcoin (BTC/USD), have historically been constrained by an epistemological bifurcation: numerical time-series models remain blind to real-world narrative catalysts and structural regime shifts, while unstructured natural language processing systems lack quantitative microstructural precision and fail under execution friction. Concurrently, decentralised prediction markets (such as Polymarket) aggregate continuous, forward-looking Arrow-Debreu state prices, yet their raw signals are routinely degraded by uninformed retail noise, Sybil wash-trading, and whale liquidity concentration.
In this paper, we introduce Dubstrata Causal-FM, an institutional-grade, bi-temporal multivariate causal foundation model specifically engineered for non-stationary, high-volatility financial time series. Causal-FM operates on a continuous uniform 5-minute canonical grid (N = 210,240 timesteps over a 24-month horizon, comprising 3,363,840 data points) organised into a modular 16-Channel Multivariate Tensor across four epistemological quadrants.
The neural architecture employs a modified Google TimesFM-3 backbone featuring decoupled 2D Grid Attention (causal temporal horizontal attention across a 512-step lookback window paired with variate vertical cross-attention across the 16 channels), Reversible Instance Normalisation (RevIN), and a 9-quantile continuous loss head (q0.10 through q0.90) regularised with a context-cutoff referenced delta directional hinge loss and guaranteed monotonic sorting.
1. The Epistemological Divide in High-Volatility Asset Forecasting
Forecasting high-volatility asset dynamics—such as BTC/USD—presents unique econometric hurdles. First, crypto markets operate without circuit breakers, trading halts, or central exchange opening bells. Second, they display extreme retail and institutional participation across globally fragmented venues. Third, they are governed by extreme reflexive leverage via perpetual swaps.
Contemporary quantitative systems suffer from an epistemological bifurcation:
Stationary Time-Series Blindspot
Traditional econometric models (ARIMA, GARCH, standard Transformers) extrapolate historical price and volatility patterns but remain entirely blind to unfolding macroeconomic catalysts, regulatory indictments, and sudden structural regime shifts.
LLM Momentum Extrapolation & Fee Burn
Frontier LLMs process unstructured text but lack microstructural precision. When prompted for directional asset predictions, they suffer from extreme momentum extrapolation (buying local tops), generate uncalibrated convictions, and destroy capital under execution friction.
Decentralised prediction markets (such as Polymarket) aggregate continuous, incentive-aligned Arrow-Debreu state prices that reflect forward-looking event probabilities. However, raw prediction market feeds are polluted by thin-market illiquidity, coordinated Sybil wash-trading, and retail noise herding. Dubstrata Causal-FM resolves these failure modes through rigorous on-chain graph forensics and bi-temporal conditioning.
2. The Master 16-Channel Multivariate Epistemological Tensor
Causal-FM structures market reality into a canonical uniform 5-minute grid across four distinct quadrants, fusing physical market microstructure with prediction contract forensics and causal knowledge diffusion:
3. Neural Architecture: Modified TimesFM-3 Backbone & 2D Grid Attention
Standard time-series transformers either flatten multivariate channels into a single 1D sequence (exploding computational complexity to O(L² · M²)) or use channel-independent projections (ignoring cross-market causal transmission). Causal-FM resolves this with decoupled 2D Grid Attention across a lookback window of L = 512 steps (42.67 hours of continuous 5-minute data):
Each transformer layer alternates between two orthogonal attention operations:
The model utilizes Reversible Instance Normalisation (RevIN) to handle non-stationarity without losing volatility amplitude, paired with a 9-Quantile Continuous Headevaluated across q ∈ {0.10, 0.20, 0.30, 0.40, 0.50, 0.60, 0.70, 0.80, 0.90}. To eliminate quantile crossing (where a higher quantile falls below a lower quantile), the loss function enforces monotonic sorting and delta directional hinge regularization.
4. Certified Out-of-Sample Forward Performance Scorecard
The model was evaluated out-of-sample across 149 rolling hourly test windows (September 15 to September 22, 2026) directly on the production host connected to live market feeds. The scorecard below documents performance across all six operational horizons:
| Horizon Target | 5m (1 step) | 15m (3 steps) | 1h (12 steps) | 4h (48 steps) | 8h (96 steps) | 24h (288 steps) |
|---|---|---|---|---|---|---|
| Directional Hit Rate (DHR) | 54.36% | 59.06% | 68.46% | 67.11% | 67.79% | 61.07% |
| Matthews Correlation (MCC) | +0.0754 | +0.1802 | +0.3706 | +0.3433 | +0.3589 | +0.2499 |
| Continuous CRPS Loss | 0.0009 | 0.0009 | 0.0009 | 0.0011 | 0.0015 | 0.0024 |
| Mean Pinball Loss | 0.0004 | 0.0005 | 0.0005 | 0.0006 | 0.0008 | 0.0012 |
| Return MAE (bps) | 10.26 bps | 10.61 bps | 11.01 bps | 13.39 bps | 17.51 bps | 26.41 bps |
| Variance Capture Ratio | 0.85x | 0.83x | 0.88x | 0.83x | 0.68x | 0.57x |
| Institutional Hurdle (≥60%) | Positive Edge | Near-Pass (59.1%) | PASS (68.5%) | PASS (67.1%) | PASS (67.8%) | PASS (61.1%) |
The model decisively clears the strict institutional hurdle of ≥60.00% directional accuracy across all primary macro holding horizons (1h to 24h), while maintaining an ultra-low return MAE of 10.26 bps at the 5-minute tactical execution boundary.
5. Live 5-Minute Prediction Market Harness Audit (September 25, 2026)
To stress-test Causal-FM in live, high-frequency environments, a dedicated forward audit was conducted on September 25, 2026, evaluating 135 continuous 5-minute decision cycles on Polymarket's “Bitcoin Up/Down 5-min” orderbook:
The Critical High-Frequency Insight: “Magic in the Harness”
In sub-15-minute regimes, the raw point forecast exhibits a near-coin-flip edge (54.36%). However, when combined with asymmetric conviction gating (gating trades to >18 bps expected margin) and maker-only liquidity posting, the live execution harness converted marginal raw predictions into a 67.50% win rate and 1.91x profit factor. In ultra-high-frequency regimes, harness engineering and execution filtering are more critical than raw point accuracy.
6. Multi-Model Benchmark Tournament: Causal-FM vs 4 Baselines vs 10 Frontier LLMs
To empirically prove whether generalist reasoning LLMs can compete with specialised continuous causal foundation models, we benchmarked Causal-FM against 4 classical econometric baselines and 10 frontier 2026 AI models across identical point-in-time forward test windows (172 decision cycles, 5 bps flat taker fee + Kyle non-linear slippage):
| Rank | Model System | Category | DHR (%) | MCC | Win Rate | Profit Factor | Max Drawdown |
|---|---|---|---|---|---|---|---|
| #1 | Dubstrata Causal-FM | Causal Foundation | 66.86% – 86.67% | +0.3365 – +0.7559 | 72.12% | 2.73x | 2.37% – 3.16% |
| #2 | Rolling Lookback Mean | Econometric Baseline | 64.53% | +0.3122 | 64.17% | 1.86x | 2.56% |
| #3 | VAR-16 Ridge Regression | Econometric Baseline | 60.47% | +0.2332 | 59.32% | 1.85x | 6.06% |
| #4 | Inception Mercury 2.5 | Frontier Reasoning LLM | 60.00% | 0.0000 | 60.00% | 2.23x | 0.43% |
| #5 | GBDT Multivariate (HistGradient) | Tabular ML Baseline | 49.42% | +0.0114 | 49.52% | 0.99x | 10.09% |
| #6 | Sakana Fugu Ultra v2 | Frontier LLM | 40.00% | -0.2887 | 40.00% | 1.45x | 0.85% |
| #7 | InclusionAI Ling 3.0 Flash VL | Frontier LLM | 40.00% | -0.2774 | 40.00% | 1.18x | 0.50% |
| #8 | Nex AGI Nex N2.5 Pro | Frontier LLM | 26.67% | -0.4725 | 26.67% | 0.76x | 0.85% |
| #9 | DeepSeek Pro Latest | Frontier LLM | 26.67% | -0.4264 | 26.67% | 0.65x | 0.91% |
| #10 | Meta Muse Spark 1.3 | Frontier LLM | 33.33% | 0.0000 | 33.33% | 0.00x | 0.00% |
| #11 | Qwen 3.8 Max 0902 | Frontier LLM | 13.33% | -0.3536 | 13.33% | 0.32x | 0.67% |
| #12 | OpenAI GPT Astra Latest | Frontier LLM | 20.00% | -0.5774 | 20.00% | 0.26x | 0.92% |
| #13 | Stealth Union Alpha | Frontier LLM | 20.00% | -0.5774 | 20.00% | 0.26x | 0.92% |
| #14 | DeepSeek Flash Latest | Frontier LLM | 13.33% | -0.1066 | 13.33% | 0.17x | 0.80% |
| #15 | Martingale Persistence (Random Walk) | Naive Baseline | 0.00% | 0.0000 | 0.00% | 0.00x | 0.00% |
The 6 Critical Takeaways from the Frontier AI Benchmark:
- 1. Momentum Extrapolation Fallacy:9 out of 10 frontier LLMs registered sub-martingale accuracy (13.33% – 40.00% DHR). Because web training text is hindsight-biased, LLMs extrapolated recent price surges right into turning points, repeatedly buying local tops.
- 2. Classical Econometrics Beat 9 of 10 LLMs:Simple Rolling Mean (Rank #2, 64.53% DHR) and linear VAR-16 Ridge (Rank #3, 60.47% DHR) crushed every frontier model except Inception Mercury 2.5.
- 3. GBDT Collapse:Gradient-boosted decision trees collapsed to a coin-flip (49.42% DHR) and suffered the highest drawdown (10.09%), proving that tabular ML lacks the temporal self-attention required for dynamic time series.
- 4. Absence of Calibrated Quantiles:LLMs generate qualitative textual convictions without downside Value-at-Risk boundaries. When scaled under execution fees, they rapidly destroy capital.
- 5. Latency Disqualification:Frontier LLMs exhibited 1,200ms – 4,800ms inference latencies, disqualifying them from real-time high-frequency execution gateways where sub-100ms response is mandatory.
- 6. Inception Mercury 2.5 Outlier:Inception Mercury 2.5 was the sole LLM competitor (Rank #4, 60.00% DHR, 2.23x profit factor), demonstrating that deep chain-of-thought financial models can occasionally isolate regime shifts.
7. Horizon-Resolved Attribution (HRX) & Three Institutional Faithfulness Proofs
Standard explainability tools (such as static SHAP) suffer from the static lookback invariance fallacy—assuming that a feature's importance is invariant whether forecasting 5 minutes or 24 hours ahead. Causal-FM introduces the Horizon-Resolved Attribution Engine (HRX), computing point-in-time attribution tensors across horizons H, historical lookbacks L, and channels M:
Deletion Gain Invariant
Certified Deletion Gain of ΔGain = +0.1899. Perturbing top-attributed features degrades accuracy substantially faster than random channel masks.
Horizon Shuffle Margin
Certified Horizon Shuffle Margin of +0.0442 (p < 0.001), disproving static lookback invariance and verifying true horizon-specific dynamics.
Zero-Lookahead Invariant
Gradient ratios with respect to unobserved future states are identically 0.0000 across all 149 rolling windows, satisfying Federal Reserve SR 11-7 model risk governance.
HRX demonstrates that at short horizons (5m–15m), microstructure channels (Ch 0–3) drive 93.67% of predictive power. As the forecast horizon extends to 1h–4h, prediction market forensic channels (Ch 4–7) surge, proving an empirical 30-to-60 minute predictive lead time for decentralised prediction orderbooks over spot liquidity.
8. Production Directives: Turning Alpha into Net Institutional PnL
Simulating a $1,000,000.00 institutional capital allocation across 149 forward windows demonstrated gross trading alpha of +$55,994.53 across $111M notional turnover. However, paying 5.0 bps flat taker fees consumed 100% of the raw profits. The whitepaper establishes two non-negotiable production deployment directives:
Passive Maker Order Routing
All institutional portfolio rebalances must be routed exclusively via post-only limit orders or fee-rebate maker tiers (0.0 bps fee or maker rebates), immediately recovering the underlying 2.73x – 3.53x profit factor.
Raise Conviction Threshold to ≥18.0 bps
Raising entry threshold gating from 4.0 bps to 18.0 bps eliminates 70% of low-margin churn trades while preserving 82% of gross trading profits.
@article{khuzwayo2026causalfm,
author = {Khuzwayo, Muzikayise},
title = {Dubstrata Causal-FM: A Bi-Temporal Causal Foundation Model with 2D Grid Attention,
Prediction Market Forensics, and Knowledge Graph Diffusion for High-Volatility Asset Forecasting},
journal = {SSRN Electronic Journal},
year = {2026},
month = {October},
url = {https://papers.ssrn.com/abstract=7539038},
doi = {10.2139/ssrn.7539038},
note = {Dubstrata Research Desk. SSRN ID 7539038}
}