Most quantitative forecasting systems evaluate accuracy in a vacuum. They ask a single post-hoc question: “How close was the prediction to the price?”
At Dubstrata, we believe raw price proximity is an incomplete—and often misleading—measure of model intelligence. In non-stationary financial markets, a model can be directionally right for the wrong causal reasons. If an online learning loop reinforces a lucky coincidence, it degrades out-of-sample reliability.
Today, we deployed our Convergent Multi-Horizon Accuracy Emitter (wss://api.dubstrata.com/api/v1/ws/accuracy) and Causally-Gated Low-Rank Adaptation (CG-LoRA) pipeline on top of our 16-channel TimesFM-3 foundation architecture.
Instead of chaining short-term predictions, our engine issues simultaneous, independent forecasts across seven distinct horizons (5m, 15m, 30m, 1h, 4h, 8h, 24h) at every 5-minute cycle, each conditioned strictly on the information filtration available at issuance (T − h). When settlement time T arrives, all seven historical vintages converge onto a single present timestamp and are audited simultaneously against the realized oracle price.
Six Formal Accuracy & Diagnostic Measures
To turn real-time evaluation into continuous model improvement, every convergent cohort is scored across six formal accuracy and diagnostic measures:
Raw Basis-Point Proximity (Abs Error bps): Measures absolute dollar and basis-point distance between each vintage's median forecast (q50) and realized spot price at settlement T. This establishes raw execution proximity across the term structure.
Horizon-Normalized Skill Score (Snorm): Raw error naturally favours a 5-minute forecast over a 4-hour forecast simply because spot price has had less time to diffuse. Snorm benchmarks each horizon's error against a volatility-scaled random walk (σ5m√(h/5)). A 20 bps error over a 4-hour macro move represents substantially higher predictive skill than a 9 bps error over a quiet 5-minute bar.
Causal Attribution Veracity (CAV): At issuance (T − h), our HRX Jacobian explainability engine records what the model believed would move price across four quadrants:
FracDiff Price, Garman-Klass Volatility, OFI × VPIN, Funding Skew
Polymarket/Kalshi Consensus, Conviction Entropy, Wash-Trade Haircut, Capital Imbalance
Friedkin-Johnsen Belief Equilibrium, Financial Impact Flow, Embedding Drift, Conflict Entropy
On-Chain Solvency, Primary Equity Correlate, Liquidity PCA, Event Clock
At settlement T, CAV ∈ [0, 1] computes the cosine similarity between those ex-ante attribution weights and the ex-post realized shocks across all 16 channels between T − h and T.
The 4-Regime Diagnostic Matrix: By crossing forecast accuracy with Causal Attribution Veracity (CAV), every resolved forecast is classified into one of four regimes:
Accurate forecast driven by the exact factor that shocked the state space ex-post.
Accurate directional call, but the model's attributed driver remained flat while an unrelated factor moved price.
Correct causal driver identified ex-ante, but magnitude scaling under- or over-shot realized drift.
Forecast miss caused by a shock in an under-weighted quadrant.
Directional Hit Rate (DHR) & Continuous Zero-Agnostic Return (CZAR) Alignment: Tracks directional sign precision while penalising sign reversals asymmetrically relative to magnitude overshoots, preventing the model from collapsing toward a zero-return mean prediction.
Empirical Quantile Cone Coverage (q10–q90 Calibration): Audits whether realized settlement prices remain calibrated within the predicted 9-quantile uncertainty distribution across volatility regimes.
Closing the Loop: Causally-Gated Multi-Horizon LoRA (CG-LoRA)
While our base TimesFM-3 foundation backbone remains strictly frozen to prevent catastrophic forgetting, these six telemetry measures feed directly into our 38,400-parameter Low-Rank Adaptation (LoRA) layers every 5 minutes:
• When a vintage forecast (for example, a 30m or 4h forecast) resolves as TRUE_CAUSAL_ALPHA, its historical 16-channel feature state at T − h and its specific horizon projection slot receive up to 3.0x gradient importance weighting.
• When an instance resolves as REGIME_BLINDSPOT, a causal prior alignment loss shifts our cross-variate attention priors toward the quadrant that actually drove price.
• Crucially, when an instance resolves as LUCKY_NOISE_HIT, its gradient weight is damped to 0.20x—ensuring the model never corrupts its weights by reinforcing a lucky guess.
Measuring accuracy is no longer a passive back-office report. It is the live causal filter that governs how our foundation model adapts to reality.