The QuantFactor Engine — How NeuralQuant Selects Stocks
Quantitative research, made transparent. Every score is derived from audited financials, live price data, and peer-relative metrics — not opinion.
IRS% Scoring System
The Investment Readiness Score combines growth quality, valuation discipline, and risk efficiency into a single 0–100% metric.
G Score
Range: -12 to +12Measures Growth + Return + Valuation alignment. Positive means the company grows profitably without overvaluation.
- Revenue growth vs sector median
- ROE and ROIC sustainability
- P/E, P/B, and EV/EBITDA relative to peers
- Free cash flow conversion
Risk Efficiency Score
Range: -8 to +8Captures volatility, leverage, and drawdown risk. Q4 is the sweet spot — low enough risk, high enough return.
- Beta vs benchmark (lower is better)
- Debt-to-equity and interest coverage
- Max drawdown over trailing 12 months
- Altman Z-score distress filter
IRS% Composite
Range: 0% to 100%Weighted composite of G Score and Risk Efficiency, normalized to a percentile rank across the full universe.
- 60% G Score weighting
- 40% Risk Efficiency weighting
- Sector-relative normalization
- Rebalanced nightly on fresh data
IRS% Derivation Formula
G Score range (24) / Total range (40) = 60%. Risk Efficiency range (16) / Total range (40) = 40%. The additive formula naturally produces a 60/40 weight split.
All quintile scoring is computed within index cohorts: SP500 stocks score against SP500 peers; SP400/SP600 against their own groups. This eliminates size bias — a SmallCap stock is never penalized for being small.
IRS% Quintile Zones
| IRS% Band | Research band | Typical Profile |
|---|---|---|
| 80 – 100% | Upper band (top) | High ROE, low beta, reasonable valuation, consistent FCF |
| 65 – 79% | Upper band | Solid growth with manageable risk; sector leader or challenger |
| 45 – 64% | Middle band | One factor is weak (e.g., high P/E or elevated leverage) |
| 25 – 44% | Lower band | Multiple red flags: low growth, high debt, or poor returns |
| 0 – 24% | Lower band (bottom) | Distress signals, negative earnings, or extreme volatility |
PARA-DEBATE Engine
Six specialist AI agents analyze each stock from independent perspectives. A seventh — the Head Analyst — synthesizes their arguments into a single investment verdict.
Fundamental
20%Audited financials, ROE sustainability, P/E and P/B peer comparison, free cash flow conversion.
Technical
16%Price momentum across 3M/6M/1Y/2Y windows, support/resistance levels, relative strength.
Sentiment
12%News sentiment scoring, insider transaction signals, short interest, earnings surprise direction.
Macro
12%Interest rate trajectory, CPI trends, sector rotation within the economic cycle.
Adversarial
20%Stress-tests bullish assumptions. Identifies hidden leverage, governance risks, and bubble signals. Must output BULL only when data overwhelmingly supports it.
Geopolitical
12%Regulatory actions, trade sanctions exposure, supply-chain concentration, ESG compliance risks.
Consensus Mechanism
How the 7 agents reach a verdictStep 1 — Parallel Analysis
All 6 specialist + adversarial agents run in parallel, each producing a stance (BULL / NEUTRAL / BEAR) and a conviction level (HIGH / MEDIUM / LOW).
Step 2 — Conviction-Weighted Consensus
Step 3 — Verdict Guidance
| Consensus | Allowed band |
|---|---|
| > +0.5 | Upper band |
| +0.25 to +0.5 | Upper or middle band |
| -0.25 to +0.25 | Middle band |
| -0.5 to -0.25 | Lower or middle band |
| < -0.5 | Lower band |
Step 4 — Head Analyst Synthesis
The Head Analyst receives all 6 agent outputs + verified raw data. It synthesizes a final research band (upper / middle / lower) with investment thesis, bull/bear cases, and risk factors. The band is clamped — cannot deviate more than 1 tier from consensus guidance.
- •Metric hallucination scan — all agent claims validated against verified data (±30% tolerance for agents, ±15% for Head Analyst)
- •Severe fundamental flags (e.g. negative ROE) algorithmically override FUNDAMENTAL agent to BEAR
- •Head Analyst must explicitly address the Adversarial agent's challenges — cannot dismiss without rebuttal
Market Regime Overlay 8% weight
A Hidden Markov Model classifies the current market into one of four regimes — Risk-On, Late-Cycle, Bear, Recovery. This regime label is injected into every agent's context and weighted 8% in the Head Analyst's synthesis. In Bear regimes, the overlay pushes verdicts conservative; in Risk-On regimes, it allows more aggressive positioning.
Selection Logic
The engine narrows 1000+ tickers down to a focused watchlist through a series of deterministic filters.
Three Pools
Stocks are segmented into LM250 (Large + Mid), SmallCap250, and MicroCap250. Each pool has its own sector median baseline and volatility expectation.
Sell Thresholds
Automatic exclusion triggers: G Score < -4 (deep value trap or distressed), Risk Score < -3.5 (excessive leverage or volatility).
Neutral Category
Stocks with G Score < -0.5 are flagged Neutral. They remain in the universe but are deprioritized unless Risk Score is exceptionally strong (> +5).
Sector Exclusions
Mining & Metals are excluded from primary recommendations due to commodity cyclicality and unpredictable regulatory intervention.
Selection Pipeline
Data Sources
Every score is only as good as the data behind it. We source from established providers with audited track records.
Financial Modeling Prep
Premium fundamentals, ratios, key metrics, and income statements.
yfinance
Real-time prices, historical OHLCV, and dividend history.
NSE India
Official Indian equity data, corporate actions, and listings.
Finnhub
Insider transactions, news sentiment, and earnings calendars.
Data Pipeline Architecture
FMP feeds fundamentals (profile, ratios, income statements) → yfinance supplies live prices & historical OHLCV → Finnhub adds insider/sentiment → NSE provides Indian equity data. All data is refreshed nightly at 02:00 UTC.
Model Governance & Updates
IRS% recomputed at 02:00 UTC with latest fundamentals and prices
Index constituents refreshed every quarter; new IPOs/graduations added
G Score 60% / Risk 40% weights are not overfitted to any period
All PARA-DEBATE agent prompts are in git; changes are tracked and auditable
Limitations & Risks
No quantitative model is perfect. Here is what the engine does not capture — and why live results may differ from backtests.
Survivorship Bias
Backtests run on today's constituents. Delisted or merged companies are excluded, which may inflate historical performance.
Look-Ahead Bias
We mitigate this by using only data available at the rebalance date. However, restated financials can still introduce subtle bias.
Transaction Costs
Backtests assume zero slippage and no brokerage. In live trading, stamp duty, STT, spreads, and impact costs reduce realized returns.
Liquidity Assumption
Micro-cap picks may not absorb large capital deployment. The model does not model market impact for position sizing.
Regime Change
Models trained on bull-market data often underperform in bear markets.
Black Swan Events
Geopolitical shocks, currency crises, or pandemics are inherently unpredictable and not priced into historical backtests.
Position Sizing & Stop-Loss
The IRS% system does not prescribe position sizing or stop-loss levels. Portfolio construction and risk management are the investor’s responsibility. Equal-weight backtests assume no capital constraints.
Drawdown Limits
There is no maximum drawdown circuit-breaker in the scoring model. A stock with IRS% > 65 can still decline significantly if fundamentals deteriorate between rebalance dates.
SEBI Disclaimer
Past performance does not guarantee future results. These are backtested results on historical data. NeuralQuant is a research tool, not a SEBI-registered Investment Advisor, Portfolio Manager, or Research Analyst. Nothing on this page constitutes investment advice. Please consult a SEBI-registered financial advisor before making any investment decisions.
Ready to explore the engine?
Run your own screens and ask Morgan anything about your portfolio.