QuantHive-Research-Tech/horizon-v1

Horizon v1 — Causal Financial World Model

The first causal world model for financial markets achieving Pearl Level 3 (counterfactual reasoning).

What This Model Does

Generate realistic multi-asset market scenarios from natural language descriptions:

from horizon.inference.scenario_engine import ScenarioEngine

engine = ScenarioEngine.from_checkpoint("model.pt")
result = engine.generate(
    description="RBI cuts repo rate by 50bp amid slowing growth",
    instruments=["NIFTY50", "BANKNIFTY", "HDFCBANK", "INFY", "TATASTEEL"],
    n_scenarios=1000,
    horizon_days=21,
)
# result.paths: (1000, 21, 5) daily log-returns
# result.prices: (1000, 22, 5) price paths
# result.var_95, result.expected_shortfall_95

Architecture

• Base: Diffusion Transformer (DiT-1D), 172M parameters
  • d_model=768, 16 axial attention blocks, 12 heads
  • Alternating time-axis / asset-axis attention
  • adaLN-Zero conditioning + cross-attention to event tokens
• Causal Layer: Backdoor-adjusted guidance with explicit causal graph
  • 14 market nodes, 18 directed edges
  • Per-channel guidance masking (prevents placebo bleed)
  • Empirical calibration from 42 RBI rate events
• Counterfactual: DDIM inversion (100-step, 0.975 roundtrip correlation)
• NL Interface: 3-tier parser (regex → LLM → heuristic)

Key Results

CausalFinBench (Novel Benchmark)

Tier	Tests	Result
A: World Properties	5/5	Consistency, Asymmetry, Compositionality, CF Coherence, Robustness
B: Causal Validity	3/3	Placebo 100%, Real effects 100%, Sensitivity monotonic
C: Natural Experiments	1/1	42/42 RBI rate decisions — 100% direction match

Calibration

Intervention	Model Effect	Historical Actual
RBI rate cut	+0.40%	+0.41% (2.4% error)
RBI rate hike	-0.31%	-0.35%
India VIX spike	-0.72%	—
FII selling	-0.49%	—
Global risk-off	-0.44%	—

Pearl's Causal Ladder

• Level 1 (Association): Base DiT generates statistically valid paths
• Level 2 (Intervention): do-operator via causal graph surgery + guided sampling
• Level 3 (Counterfactual): DDIM inversion + action + prediction (verified on June 4, 2024 election)

Training Data

• 50 Nifty50 constituents × 26 years (2000-2026)
• 1.05M real news headlines from QLake
• 48 macro series (repo rate, CPI, VIX, yields, commodities, FX)
• 317,650 training windows

World Model Properties (5/5 PASS)

1. Consistency: Same intervention + same noise → identical output
2. Causal Asymmetry: do(rate_cut) ≠ do(nifty_rally) — different mechanisms (corr=0.47)
3. Compositionality: Combined interventions produce combined effects (56% of linear sum)
4. Counterfactual Coherence: Inversion roundtrip at 0.975 correlation
5. Robustness: Extreme interventions (±10σ) remain finite and reasonable

Limitations

• 5 instruments per call (can generate correlated paths for any 5 Nifty50 stocks)
• Causal graph is hand-specified (not learned from data)
• Magnitude calibration fitted from historical events — may not generalize to unprecedented scenarios
• NL understanding via external parsing — model itself does not natively understand text (v2 will fix this)
• No options/derivatives pricing (equity paths only)

Usage Requirements

• PyTorch >= 2.0
• ~700MB disk for checkpoint
• GPU recommended for inference (<30s on A100 for 1000 paths)
• CPU inference: ~20 minutes for 1000 paths

Citation

@software{horizon_v1_2026,
  title={Horizon: A Causal Financial World Model for Indian Markets},
  author={QuantHive Research},
  year={2026},
  url={https://huggingface.co/QuantHive-Research-Tech/horizon-v1}
}

License

Apache 2.0