Prediction Model Overview

All quantitative prediction systems in MI-OS XUANJI | 28 engines across 7 categories

Prediction Engines

MC Factor Models

HIGH Tier

2,579

Total Predictions

41K+

mc_history Data Pts

MC Factor Model

SCPM / Supply Chain

Geopolitical

Credit / Risk

Housing / Regional

Valuation / DCF

1. MC Factor Models (18 models)

Ridge regression on 20yr monthly data + Newey-West HAC + residual bootstrap 20K simulations. FX Dashboard

Methodology: OOS R2 = rolling 60/40 out-of-sample validation. Factor stability = coefficient sign must not flip between data halves. See docs/MC_Prediction_Engine.md for full causal maps + failed experiments.

2. SCPM — Supply Chain Propagation Model

Per-industry 3-variable cycle model: L1 (Leading) + O (Operating) + R (Risk). 28+ industry configs.

SCPM Cycle Phase ACTIVE

Composite score [-6, +6] maps to 6 phases: TROUGH / EARLY_RECOVERY / MID_CYCLE / LATE_CYCLE / PEAK / DOWNTURN

L1 (Physical Lag, 3-9M lead) SOX momentum + monthly revenue YoY + FRED orders O (Order/Inventory) Inventory days z-score + pass-through ratio + capex/depreciation R (Risk/Regressor) HY spread + VIX + yield curve + NFCI

28+ industriesA/B/W/C data tiersCalibrated per-industry thresholds

SCPM Z-Score Predictions ACTIVE

Statistical anomaly detection: when any SCPM indicator exceeds 2-sigma → generate directional prediction

Indicator z-score > 2.0 → prediction (direction + magnitude from historical distribution)

Per-tickerQuantitative basis

SC Contagion Engine ACTIVE

Supply chain risk propagation. N-hop decay model through verified SC edges.

Score shock at Node A → 50% decay per hop (regime-coupled) → STAGFLATION: 35% decay (wider) → GOLDILOCKS: 55% decay (contained) → Max 4 hops

936 SC edgesRegime-coupled decay

SC Lead-Lag Prediction ACTIVE

Upstream revenue changes predict downstream margins (3-6M lag). Based on verified supply chain relationships.

Upstream RevG change → 3-6M → Downstream margin impact Dampening: upstream=0.6, downstream=0.4 (auto-tuned)

Per-template MAE guard8 tunable parameters

3. Macro Regime & Path Detection

4-regime classification + Bayesian 4-path probability + 18-year cycle + dollar cycle.

Macro Regime Classifier ACTIVE

4 regimes: GOLDILOCKS / REFLATION / STAGFLATION / DEFLATION. SFI (Systemic Fragility Index) = VIX(30%) + HY(30%) + YC(20%) + ISM(20%)

Leading indicators → SFI → Regime classification Growth signal (ISM, employment) × Inflation signal (CPI, PPI, breakeven) → 2×2 matrix → regime

macro_regime.pyReal-time

Bayesian Path Detection (10 signals) ACTIVE

4 paths: A (Goldilocks) / B (Reflation) / C (Stagflation) / D (Hard Landing). Bayesian posterior updating.

Prior P(path) × L(data|path) for 10 signals: S1: Yield curve S2: Claims S3: HY spread S4: Breakeven S5: NFCI S6: Bank lending S7: Dual YC inversion S8: Exogenous shocks (oil/tariff/geo/SC) S9: Dollar cycle 18yr phase + dedollarization → Normalize → Dominant path + confidence

detect_paths()10 Bayesian signals

Dollar Cycle (18yr + Chokepoint) ACTIVE

Patel 18yr cycle + He Si-Yin chokepoint framework. Current: Year 16/18, WEAKENING phase.

18yr Phase: RECOVERY→EXPANSION→BOOM→FRENZY→CRISIS Dollar Phase: EARLY_STRENGTH→PEAK→WEAKENING→TROUGH Dedollarization: 5-layer monitor (SWIFT/reserves/gold/CNY/BRICS) Chokepoint Risk: SWIFT dependency + USD exposure + sanctions proximity

dollar_cycle.py4 sub-modules

Macro Prediction Arbiter ACTIVE

Cross-module arbitration: normalizes tickers, removes stale/buggy predictions, picks canonical view when modules disagree.

macro_prediction_arbiter.pyBronze→Silver→Gold

4. Geopolitical Models

Decision trees + path forecasts + constraint-based analysis. Bridges qualitative → quantitative.

Geo Decision Trees ACTIVE

Per-issue branching with probability-weighted outcomes. E.g., GEO-002 (Taiwan Strait): 3 branches with per-ticker quantitative impact.

Issue state → Branches (3-5 paths) Each branch: probability + per-ticker direction/range + mechanism → unified_context.derived_impacts → ticker-level predictions

geo_fill_decision_trees.py15+ active issues

Geo Path × Macro Matrix ACTIVE

Macro implications by quarter per geopolitical path. Conditional probability evolution over time.

GEO issue × Path → Fed / GDP / CPI / Oil / FX targets by quarter Event checkpoints with probability updates

geo_path_forecast.pySQLite Silver/Gold

Geo Asset Allocation ACTIVE

MC-calibrated BUY/HOLD/SELL for SPX/TW50/GLD/TLT. Rule-based (not LLM).

MC model expected return + CI → P(>8%) → BUY, P(<-8%) → SELL, else HOLD + Regime override (STAGFLATION → force HOLD equities)

geo_asset_models.py4 assets

Broadmarket Price Verification ACTIVE

L1 price hook batch verification: predicted vs actual market data → market_regime_signal (risk_off/risk_on/stagflation).

geo_verify.pyReal-time verification

5. Valuation & Financial Models

DCF 3-stage + Reverse DCF + Sensitivity + Monte Carlo simulation.

DCF 3-Stage Valuation ACTIVE

calc_engine.dcf_3stage(): Explicit forecast → fade → terminal. Template-aware WACC.

FCF base × g1 (5yr) × g2 (fade) × g3 (terminal) ÷ WACC → NPV → per-share fair value Gap > 15% → prediction candidate

calc_engine.py663 cached valuations

Reverse DCF ACTIVE

Implied growth rate from market price. Identifies what the market is "pricing in".

sensitivity_engine.pyMispricing detection

MC Scenario Simulation ACTIVE

10K paths for revenue, score, portfolio returns, SC contagion. Percentile distribution output.

Revenue paths: mean-reverting growth noise (3yr) Score evolution: probabilistic changes Portfolio: multi-asset correlated scenarios SC contagion: propagation probability

xj_monte_carlo.pyp10/p25/p50/p75/p90

Capital Cycle Stage ACTIVE

5-stage detection: CC1 (Recovery) → CC2 (Expansion) → CC3 (Peak) → CC4 (Downturn) → CC5 (Trough)

RevG + Capex/Deprec + P/B + Utilization → Stage Eligible: Cyclical/Energy/Semi/Industrial/Hardware templates

cc_stage_detector.pyPortfolio Model B gate

6. Credit, Housing & Commodity

CRS credit scoring + housing regional forecast + commodity cobweb models.

CRS Credit Risk Score ACTIVE

7-dimension score: ICR(25%) + Leverage(25%) + Current(20%) + CCC(15%) + AltmanZ(15%) + Profitability + Scale

Financial data → 7 piecewise-linear subscores → Weighted composite → Rating (AAA to CC) → TCRI 1-9+D mapping for bank comparability

crs_engine.pyS&P/Moody's calibrated

Housing Price Forecast v2.4 ACTIVE

22 regions × 3 scenarios × 20 quarters. 6 models: Credit cycle + Stock-flow + Affordability + Demographic + TSMC multiplier + Policy regime.

Credit cycle (Mian-Sufi + Patel feedback) Stock-flow (permits → supply → 2026 交屋潮) Affordability constraint (PIR ceiling) Demographic (household formation) TSMC multiplier (fab investment shock) Policy regime (CBC + 囤房稅 + 新青安)

housing_forecast.py22 regions, 5yr horizon

Commodity Driver Model ACTIVE

Cobweb cycle + cost pass-through for commodity-driven companies. Edible oil, steel, energy, metals, shipping.

Commodity futures price → Cobweb production lag (1-3Q) → Cost pass-through → Margin prediction → Revenue/Earnings impact by ticker

commodity_drivers.py6 commodity categories

FX Constraint Models ACTIVE

Impossible Trinity framework for managed currencies: CNY (PBOC band), HKD (peg), SGD (NEER basket).

CNY: rate_diff + VIX + oil + DXY → depreciation signals Hard ceiling 7.5 / floor 6.5 / daily ±2% HKD: Pegged 7.75-7.85 → HKD/TWD ≈ USD/TWD SGD: MAS NEER target → partially predictable

fx_constraint_models.py3 currencies

7. Meta & Quality Systems

Divergence detection, game theory, thinking models, prediction quality gates.

Divergence Engine (12 types) ACTIVE

Detects disagreements between XUANJI signals and market consensus. Smart/Dumb money classification.

D1: Earnings vs SC D2: Insider vs Score D3: SCPM vs Fund Flow D4: Module introspection D5: Analyst vs Score D6: Macro vs Sector D7: Sentiment vs Score D8: Contagion vs Price D9: Smart vs Dumb D10: 18yr vs Consensus D11: Dollar structural D12: Dedollarization

divergence_engine.py12 divergence types

Game Theory Engine ACTIVE

3-agent competitive simulation: Client strategist vs Primary competitor vs Regulatory environment. Response: MATCH / UNDERCUT / DIFFERENTIATE / IGNORE.

game_theory_engine.pyConsulting reports

Thinking Model Auto-Learning ACTIVE

Extract reusable frameworks from articles → A/B test via predictions → promote if +5% hit rate.

CANDIDATE → VALIDATED → PROMOTED → RETIRED 43 active thinking models, 118 books integrated

thinking_model.py43 TMs active

Prediction Quality Gates (N19-N29) ACTIVE

N19: quality gate (universe + direction + deadline). N19o: rubber-stamp detection. N19p: NOWCAST/FORECAST classification. N19q: narrative vs model-backed.

xj_predict.py11 gates

MC Factor Model Detail

Full methodology: docs/MC_Prediction_Engine.md