FIN uses three different types of intelligence working together: a Large Language Model (LLM) to reason and explain, Machine Learning (ML) to spot patterns and produce precise prices, and statistics to keep everything grounded in numbers.

---

Large Language Model (LLM)

The LLM is the "brain" that produces the thesis — the qualitative reasoning behind every prediction. Every time FIN runs, it feeds a rich set of data to the LLM so it can understand the full picture:

• Stock prices — open, high, low, close, and volume with technical indicators
• News articles — scored for sentiment (positive/negative) toward the company and the economy
• Surprise signals — unexpected events like earnings beats or economic shocks
• Psychology scores — 12 cognitive bias measurements (FOMO, herding, loss aversion, etc.)
• Competitor data — peer price performance, valuation metrics, and competitor news sentiment
• Financial statements — income, balance sheet, and cash flow data
• Calendar events — earnings dates, dividend dates, and other market-moving events
• Market context — beta, correlation, and relative strength vs SPY and sector ETFs
• Options data — put/call ratios and implied volatility
• Rates & FX context — USD/EUR sovereign yields (2Y/5Y/10Y) and EUR/USD exchange rate

All of this is packed into a single briefing and sent to the LLM. The model weighs every signal, considers how they interact, and produces:

1. A qualitative thesis — direction, conviction, regime assessment, bull/bear scenarios, and falsifiable hypotheses
2. Falsifiable hypotheses — testable predictions registered as thesis monitors (checked daily against live data). Each prediction produces 2–4 hypotheses; the ones that can't be tracked or are already breached are rejected, duplicates are removed, and the rest are kept. Every day, active monitors are checked against live market data — if one is breached, it's marked as triggered and factored into subsequent predictions.
3. A baseline price — the AI's direct numeric estimate for each horizon offset.

Three numeric voices, not one

Asking an AI for an exact price in one shot is noisy — the same briefing can produce slightly different numbers on re-runs. Rather than trusting a single number, FIN asks two follow-up questions and blends all three answers, each weighted by how confident it is:

• Baseline price — the direct estimate (the original behavior, kept as-is).
• Residual correction — "Given what price history alone suggests, what small % fix does the thesis justify?" Bounded to a few percent so it nudges rather than hijacks.
• Agreement check — the AI is asked twice more with slightly more creative freedom; the average of all three becomes an extra view, and wide spread between them automatically counts as uncertainty (less weight, wider band).

If any follow-up fails, the forecast simply proceeds without that slice — nothing breaks. The lightweight intra-day updates skip both extras to stay fast; final predictions run them for every ticker.

Direction classifier

Alongside the quantile price models, FIN trains a direction classifier — a binary model that outputs the probability the price moves up over the horizon (P(up)). It is distinct from the price-estimate model and appears as its own row in the Model Reconciliation panel. When the classifier opposes the combined AI/price-model consensus, that counts as a disagreement: the confidence band is widened and confidence is trimmed, scaled to the classifier's conviction.

What Makes the LLM Approach Unique

Unlike a black-box algorithm, the LLM explains its reasoning. You can read the full thesis text to understand exactly why the system thinks the stock will move in a certain direction — which news articles influenced it, which psychology biases were most relevant, and how the technical signals factored in.

---

Machine Learning (ML)

While the LLM handles reasoning, ML produces the numeric price predictions by learning from past results.

Quantile Models

The system trains quantile GBM (gradient boosted machines) models — five models per ticker and horizon that predict different points on the return distribution:

Model	Predicts	Meaning
q10	10th percentile	"What's the worst-case scenario (90% chance returns will be better)"
q25	25th percentile	"What's the below-average scenario"
q50	50th percentile (median)	"What's the typical expected return"
q75	75th percentile	"What's the above-average scenario"
q90	90th percentile	"What's the best-case scenario (90% chance returns will be worse)"

The q10–q90 gap forms the confidence band. These bands are optionally calibrated by conformal prediction to guarantee ~80% coverage.

GBM is like asking 5 "experts" (decision trees) to work together, where each new expert focuses on fixing the mistakes of the previous ones.

Why This Matters

Instead of getting a single prediction like "stock will go up 5%", you get:

"There's a 10% chance it drops more than -2%, a 50% chance it's around +1.5%, and a 10% chance it goes up more than +8%"

How ML Learns

Every prediction run saves a feature snapshot — a complete record of everything that went into that prediction:

• Technical indicators, sentiment scores, psychology biases
• Financial fundamentals, market context, options data
• Rates & FX context (USD/EUR sovereign yields at 2Y/5Y/10Y, EUR/USD exchange rate)
• Thesis features (conviction, regime probabilities, etc.)
• How far the AI's forecast diverged from the model's — that disagreement is later used to improve the model's error-correction, so predictions get better at recognising when the two methods clash

When the target date passes, the accuracy tracker resolves the snapshot with the actual outcome. Once enough resolved snapshots accumulate (100+), the system trains new models.

Quantile Regression

Instead of predicting just the average return (which everyone does), they predict ranges:

• This tells you not just "what will happen" but "what's the range of possibilities"

Model Training Settings

Setting	What it means
Tree depth	Each "expert" tree can make ~31 different decisions
Learning rate	5% — how much each new expert adjusts the previous ones' mistakes
Feature sampling	Each tree only looks at 80% of features (prevents over-reliance on any one feature)

Tiered Architecture

At prediction time, the system picks the best available model:

1. Tier 1 — Your personal model (best accuracy, needs ~100 snapshots)
2. Tier 2 — Pooled model from all users (good fallback)
3. Tier 3 — Sector-transfer model (from a similar company)
4. Tier 4 — Thesis heuristic (always available, no training needed). Neutral theses with moderate or higher conviction still contribute a small symmetric signal, so the AI's data-quality read stays in the picture rather than dropping out silently.

Each tier is performance-gated: if its recent accuracy or coverage drops below a threshold, it's automatically downgraded. You never see a bad model's output.

Self-Correction

If a model gets stale as market conditions change, it's retired automatically. The system falls back to Tier 4 (thesis heuristic) until fresh models can be trained. The model's weight in the blend also scales with its win rate — a badly-losing model is near-excluded rather than dominating.

Cold Start

New tickers don't wait weeks. The system backfills historical feature snapshots from existing price data, enabling Tier 1 training on the first retrain cycle. Backfilled samples are weighted lower than live samples during training so the model prioritizes real outcomes over simulated history.

---

Statistics

Statistics are the foundation that keeps the LLM and ML honest. They measure, validate, and quantify everything the system does.

Prediction Accuracy

Every prediction is measured against what actually happened:

Metric	What It Tells You
Error %	How far off was the prediction? (absolute percentage difference)
Band Hit	Did the actual price fall within the confidence range? (Yes/No)
MAPE	Mean Absolute Percentage Error — the average error across recent predictions

Gating

If the system's recent MAPE exceeds a configurable threshold, predictions are automatically withheld (gated) until accuracy recovers. This prevents unreliable predictions from reaching the dashboard.

Sentiment Thresholds

News sentiment scores use statistical boundaries:

• Below 45 — negative sentiment
• 45–55 — neutral
• Above 55 — positive sentiment

Technical Indicators

Statistics power the technical analysis:

• RSI (Relative Strength Index) — a 0–100 score. Above 70 suggests overbought, below 30 suggests oversold
• Trend detection — statistical tests distinguish mean-reverting from random-walk behaviour (but never assert a trend), a separate state model classifies the regime (bullish/bearish trend, mean-reverting, high-volatility), and trend-following indicators (ADX, Donchian channel, moving-average alignment, the Hurst exponent, trend slope) provide positive evidence of an ongoing trend
• Volume analysis — whether trading volume is increasing or decreasing relative to historical norms

Psychology Scoring

Each of the 12 cognitive biases is scored on a 0–100 scale using statistical analysis of article text:

• Low (0–30) — barely detectable
• Medium (30–60) — some influence present
• High (60–100) — strongly influencing the market

A Third Opinion: The Statistical Referee

Alongside the AI and the machine-learning model, FIN computes a simple, classical forecast from the stock's own price history alone — no news reading, no trained model. It's a deliberately "dumb" number that exists to keep the smarter systems honest.

How it works, in plain terms:

• FIN first checks the recent price behaviour — does the price tend to oscillate around an average (mean-reverting), or does it show no reliable pattern (a random walk)?
• Then it produces a grounded estimate of the likely move and a range around it.

What the price history shows	What the forecast expects
Random walk (no reliable pattern)	Essentially no move — at most a small drift from the recent trend
Mean-reverting (pulls back to its average)	A move back toward its ~20-day average (e.g., if the price is well above that average, the statistical view leans down)
Inconclusive	A blend of the two above

What you get from it: an expected move, a confidence range (the wider the range, the less certain), and a confidence score that grows with more price history — a stock with only three weeks of data is trusted less than one with three months. It also has built-in honesty: it needs at least ~20 days of prices to say anything, and an extreme overnight jump (more than ~15%, which usually means a stock split or a bad data point) is ignored rather than chased.

Where you see it: in the Model Reconciliation panel on a prediction's Reasoning page, inside the purple "Statistical layer" box. It shows the role badge, the detected regime, the expected move, the range, and the confidence — and whenever the statistical view takes part in a forecast, you'll see a purple slice for it in the weight bar. The residual correction described above builds directly on this number (it corrects the statistical price), so the two are related but counted separately and transparently.

The four roles it can play (most are invisible to your live numbers):

• Shadow (on by default) — the forecast is recorded quietly for research and backtesting. It never touches the numbers you actually see.
• Honest fallback baseline — if both the ML model and the AI thesis are ever unavailable (for example, an outage), this statistical estimate becomes the published price instead of a rough sentiment placeholder.
• Shrinkage prior (on by default) — gently pulls pooled forecasts toward the statistical view.
• Plausibility anchor (on by default) — if the final forecast lands far outside the statistical envelope, FIN widens the band and lowers confidence (a "that looks suspicious" flag) without moving the point estimate.
• Options-market band floor (on by default) — a separate, market-based check that is not part of the statistical forecaster. It compares the band to the move implied by the options market's 30-day volatility and enforces a hard floor — the band is widened to at least the implied-move range when the model is tighter than traders expect, and confidence is trimmed. Because it uses the options market rather than price history, it is independent of the statistical anchor and the two never conflict. \nThere is also an empirical-error floor: if the model's band is tighter than the stock's own observed error rate, it is widened to match the historical error — so a narrow band can't over-promise against past performance.

The first two are safe to leave on. The shrinkage prior and plausibility anchor are enabled by default because their influence is bounded and fail-open, but they can be disabled per ticker if needed.

---

How They Work Together

   Collected Data
          ↓
  ┌───────────────┐
  │  Statistics   │──→ Measure, validate, set thresholds (+ independent forecast)
  └───────────────┘
          ↓
  ┌───────────────┐
  │      LLM      │──→ Reason, produce thesis, explain + 3 numeric voices
  │               │   (baseline price, residual correction, agreement check)
  └───────────────┘
          ↓
  ┌───────────────┐
  │  ML (Tier 1-4)│──→ Quantile price from thesis + features
  └───────────────┘
          ↓
  Blend all views by confidence (weights shown in Reconciliation panel — only active contributors are listed)
          ↓
   Price + band + reasoning
          ↓
  (Feedback: actual outcome resolves snapshots → retrain ML)

Statistics measure the raw data and validate the outputs. The LLM uses that data to reason about what will happen and produce a qualitative thesis plus three numeric views (baseline, residual correction of the statistical forecast, and a 3-sample agreement check). ML uses the thesis features plus all historical data to produce precise quantile price predictions. Statistics also produce an independent price-history forecast — the "statistical referee" shown in the Model Reconciliation panel — that cross-checks the other two and, when enabled, actively shapes the forecast: the shrinkage prior gently pulls the pooled estimate toward the statistical view, and the plausibility anchor widens the band if the final forecast lands far outside the statistical envelope. Neutral theses are not discarded: moderate-or-higher conviction neutral theses produce a small symmetric signal that keeps the LLM leg in the blend as a near-zero-return source, so the combined forecast can still use the LLM's data-quality read. The actual outcome feeds back to resolve feature snapshots, which train the next generation of ML models.