FIN is designed to work with default configuration, but every stock is different. This page walks you through the settings you can adjust from the dashboard to get the best results for your specific situation.

---

Quick Overview — The 6 Tuning Areas

📡

RSS Feed Frequency

How often FIN checks for news — more frequent means more accurate

🛡️

Prediction Gating

When FIN skips a prediction to protect you from bad forecasts

⏳

Signal Decay

How fast old news and psychology signals lose their influence

🧠

ML Training

How many predictions before the system learns from its mistakes

⚡

Surprise Detection

Sensitivity to unexpected events in the news

📋

Prompt Content

What data the AI sees when it makes predictions

---

Understanding the Flow

Before diving into settings, here's how FIN's pipeline works and which tuning areas affect each stage:

Collect

→

Analyze

→

Predict

→

Validate

→

Learn

Stage	What Happens	Tuning Areas That Affect It
Collect	RSS feeds are checked, stock prices fetched, calendar events gathered	Feed frequency, cache settings, competitors
Analyze	News is scored for sentiment, surprises, and psychology	Surprise thresholds, decay rates
Predict	AI combines all signals into a thesis and price prediction	Prompt content, temperature, gating, ML tier
Validate	After market close, predictions are compared to actual results	MAPE thresholds, accuracy decay
Learn	Resolved snapshots train new GBM models	Training thresholds, drift detection

---

1. RSS Feed Frequency

RSS feeds if one of the channels that FIN uses to stays informed about your stock. More frequent checking means FIN reacts to news sooner, which leads to more accurate predictions.

How to Adjust

Go to Scheduler → Feed Schedules in the dashboard. Each feed has its own schedule and settings.

Recommended Patterns

Feed Type	Recommended Frequency	Why
Ticker-specific news (Google News, Yahoo Finance, etc)	Every 1–2 hours	Your stock's direct news matters most
Central bank / policy feeds (ECB, Fed)	Every 6 hours	These publish infrequently — checking more often wastes resources
Macro / market-wide news	Every 6 hours	Broad economic context, not time-critical
Reddit discussions (with buzz)	Every 6 hours	Social sentiment changes slowly; buzz scraping is heavier

Feed Settings

Setting	What It Does	Recommendation
Schedule	Cron expression for how often to check	Every 2–3h for ticker feeds, 6h for macro
Buzz	Also collect comments and engagement data	Enable for Reddit-style feeds
Buzz Comments	Number of top comments to scrape per post	3 is a good default
Max Days	Ignore articles older than this (in days)	Default 3 — lower for faster processing
Extract Text	Full article text extraction	Enable if you want deeper analysis
Active	Toggle feed on/off	Disable feeds that consistently error

💡

Tip: The biggest accuracy gains come from checking your ticker-specific feeds more often. If you only change one thing, increase the frequency of those feeds.

What to Tune Based on Your Stock

Stock Type	Strategy
Active stock with lots of news	Check ticker feeds every 1–2 hours
Quiet stock with infrequent news	Every 4–6 hours is fine
Want faster processing	Lower max_days (e.g., 2 instead of 3)
Want more context	Add more feeds, increase buzz_comments
Resource-constrained	Reduce parallelism, increase cache TTLs

---

2. Prediction Gating

When FIN's predictions have been consistently off, the system automatically skips predictions until accuracy improves. This protects you from bad forecasts.

How It Works

1. FIN tracks prediction accuracy using MAPE (Mean Absolute Percentage Error)
2. When MAPE exceeds your threshold, predictions are "gated" (skipped)
3. After a cooldown period, FIN forces a real prediction to check if accuracy has recovered
4. If still inaccurate, gating continues

Settings (Config → Prediction Config)

Setting	Default	What It Does
Prediction Gating Enabled	On	Master switch — turn off to bypass all MAPE gating
MAPE Threshold (per horizon)	8.0%	Predictions are skipped when error exceeds this
Skip Cooldown (per horizon)	5 days	After gating, force a real prediction after this many skips
Accuracy Decay Days	60 days	How far back to look when calculating MAPE

How to Tune

MAPE Threshold controls how strict gating is:

Setting	Label	When to Use
≤ 5%	Strict	Conservative — skip early, only show high-confidence predictions
≤ 10%	Moderate	Default — good balance for most stocks
≤ 15%	Relaxed	Tolerate more error — useful for volatile stocks
> 15%	Lenient	Rarely gate — useful during earnings season

Skip Cooldown controls how quickly FIN tries again after gating:

• Lower (1–3): Resume predictions sooner — good if you want to see updates even with lower accuracy
• Higher (7–14): Stay in safety mode longer — good if you only want high-confidence predictions

Accuracy Decay Days controls how fast old errors age out:

• Lower (30 days): Recover from bad spells quicker — old high-error predictions stop affecting gating sooner
• Higher (120+ days): Keep a longer memory — more stable but slower to recover from a bad period

💡

Tip: Start with defaults (8% / 5 / 60). If predictions seem too conservative, raise the MAPE threshold. If the system stays in gating too long, lower the accuracy decay days.

---

3. Signal Decay — Sentiment & Psychology

News and psychology signals lose relevance over time. Decay rates control how fast older signals lose their influence on predictions.

Why It Matters

A news article from yesterday is more relevant than one from two weeks ago. But for long-term predictions (60+ days out), even older articles carry some useful information. Decay rates let you control this balance.

Settings (Config → Decay Config)

Each horizon has two independent decay sliders:

Setting	Short (1–10 day)	Medium (100 day)	Long (300–600 day)
Sentiment Decay	0.3	0.1	0.05
Psychology Decay	0.15	0.08	0.04

What the Numbers Mean

Sentiment Decay controls how fast news articles lose weight:

• Higher (0.3–0.5): Older news fades fast — recent headlines dominate
• Lower (0.01–0.1): News retains influence longer — more memory of past events

Psychology Decay controls how fast bias scores age:

• Higher (0.15–0.3): Recent bias readings dominate (FOMO, anchoring, herding)
• Lower (0.01–0.08): Historical bias patterns stay relevant longer

When to Adjust

Situation	Adjustment
Predictions overreact to old news	Increase sentiment decay
Predictions miss recent shifts	Decrease sentiment decay
Volatile, fast-moving market	Higher decay (focus on fresh signals)
Stable, trending market	Lower decay (longer memory helps)
Short-term predictions too noisy	Increase psychology decay for short horizon
Long-term predictions missing bias patterns	Decrease psychology decay for long horizon

---

4. ML Training — Snapshots & Tiers

This is the heart of how FIN learns. Every prediction creates a feature snapshot — a complete record of everything that went into it. When the target date passes, the snapshot is resolved with the actual outcome, turning it into a training example.

What Are Feature Snapshots?

Each snapshot captures 86+ features:

Category	Examples
Technical	RSI, ATR, trend direction, volatility, SMA crossover
Sentiment	Ticker sentiment, economy sentiment, momentum, buzz
Surprise	Count, magnitude, direction of recent surprises
Psychology	12 cognitive bias scores (FOMO, anchoring, herding, etc.)
Fundamentals	P/E ratio, revenue growth, EPS growth, debt/equity
Calendar	Days to next earnings, days to next dividend
Market Context	Beta, correlation with SPY, relative strength
Options	Put/call ratio, implied volatility
Thesis	18 features: conviction, regime, bull/bear probabilities, etc.

After the target date, the snapshot is resolved with actual_return_pct — the real return from prediction date to target date. This becomes the training label.

Training Thresholds (Config → ML Training Config)

Setting	Default	What It Does
Min Training Samples	100	Minimum resolved snapshots before any model trains
Min New Samples for Retrain	20	New snapshots needed to trigger automatic retraining
Horizon Min Samples — Short	50	Minimum for short-horizon Tier 1 activation
Horizon Min Samples — Medium	40	Minimum for medium-horizon Tier 1 activation
Horizon Min Samples — Long	30	Minimum for long-horizon Tier 1 activation
Drift MAE Threshold	30%	MAE degradation that triggers a drift alert
Deactivate Underperform Days	30 days	How long a bad model survives before auto-deactivation

How Many Snapshots for Tier 1?

The system needs enough resolved snapshots to train a reliable model. Here's what to expect:

Horizon	Snapshots per Day	Days to 50 Snapshots	With Backfill
Short (1, 2, 10 day)	3	~17 days	Same day
Medium (100 day)	1	~50 days	Same day
Long (300, 600 day)	2	~25 days	Same day

💡

Tip: With backfill (which runs automatically on first retrain), the system typically generates hundreds of resolved snapshots from your existing stock price history. This means Tier 1 training usually succeeds on the very first retrain — no waiting weeks for data.

Cold Start — What Happens with a New Ticker

When you add a new ticker, here's the journey:

Day 0: Tier 4

→

First Retrain: Backfill

→

Same Day: Tier 1

Day 0 — Tier 4 (Thesis Heuristic): No training data yet, so FIN uses the AI's qualitative thesis to make rough predictions. These are directionally useful but lack the precision of trained models.

First Retrain — Backfill: The system automatically reconstructs historical feature snapshots from your existing stock price data:

• Technical features — reconstructed from stored OHLCV price history using rolling calculations
• Fundamental features — computed on-the-fly from financial statements (P/E, revenue growth, etc.)
• Calendar features — derived from stored earnings and dividend dates
• Sentiment, psychology, surprise, thesis, options — set to NULL (the GBM handles missing values natively)
• Outcomes resolved immediately — the close price at each target date is looked up from stored data, so every backfilled snapshot becomes a resolved training example

Result: Typically hundreds of resolved snapshots in one run, enabling Tier 1 training immediately.

Drift Detection

Over time, market conditions change and old models may become stale. FIN monitors this automatically:

• Drift MAE Threshold (30%): When recent prediction error exceeds this percentage above the baseline, a drift alert fires
• Deactivate Underperform Days (30): A model that consistently underperforms for this many days is automatically deactivated
• Set to 0 to disable auto-deactivation

---

5. Surprise Detection

FIN scans articles for unexpected events — earnings beats/misses, economic surprises, guidance changes, and more. You can control how sensitive this detection is.

Settings (Config → Surprise Signals)

Setting	Default	Range	What It Does
Min Ticker Relevance	15	0–100	Articles below this relevance score are skipped
Min Magnitude Threshold	20	0–100	Surprises below this magnitude are not stored

How to Tune

Setting	Lower (Permissive)	Higher (Strict)
Min Ticker Relevance	More articles analyzed — catches borderline cases	Only highly relevant articles — less noise
Min Magnitude Threshold	Minor deviations captured — more data points	Only big surprises — cleaner signal

💡

Tip: For most stocks, the defaults (15 relevance, 20 magnitude) work well. Lower the thresholds if you want more data points for ML training. Raise them if you're seeing too many false positives.

---

6. Prompt Content — What the AI Sees

You can toggle on/off every data section in the AI's briefing. All on = most informed prediction. Turn off sections to experiment or troubleshoot.

Prediction Prompt (Config → Prediction Prompt)

These control what data the AI sees when making its initial prediction:

Category	Sections	What They Provide
Sentiment	Ticker Sentiment, Economy Sentiment, Country Economy Sentiment	Positive/negative feeling from news
Reliability	Ticker Uncertainty, Economy Uncertainty, Country Uncertainty	How trustworthy each sentiment signal is
Momentum	Avg Buzz Score, Re-acceleration, Momentum Score, User Interest	How much attention and momentum the stock has
Fundamentals	Estimates Context, Financials Context	Analyst forecasts, revenue, margins, balance sheet
Events	Surprise Context, Surprise Outcomes Context	Recent unexpected events and past reactions
Psychology	Bias Context	12 cognitive bias scores weighted by horizon
Calendar	Calendar Context	Upcoming earnings, dividends, events
Market	Market Context, Competitor Context	Benchmark comparison, peer performance
Technical	OHLCV Context	Price action, RSI, ATR, trend, volume
Options	Options Context	Put/call ratios, implied volatility
Self-awareness	Accuracy Context, ML Knowledgebase	Past prediction accuracy, model data

Self-Critique Prompt (Config → Self-Critique Prompt)

After making its initial prediction, the AI reviews its own work. These control what it sees during that review:

Section	Purpose
Ticker Sentiment	Check if prediction aligns with stock-specific news
Economy Sentiment	Check if prediction aligns with economic conditions
Momentum Score	Check if prediction aligns with recent trends
Re-acceleration	Check if prediction accounts for accelerating/decelerating signals
Surprise Summary	Check if prediction accounts for recent surprises

LLM Temperature (Config → LLM Temperature)

Controls how deterministic the AI's outputs are:

Range	Label	Effect
0.0–0.2	Deterministic	Consistent, repeatable outputs — best for predictions
0.2–0.4	Precise	Slight variation — default (0.2) sits here
0.4–0.7	Balanced	More varied — useful for thesis generation
0.7–1.0	Creative	High variation — rarely needed
1.0–2.0	Experimental	Very random — not recommended for predictions

💡

Tip: Keep temperature at 0.2 for predictions. Only increase it if you want more diverse thesis reasoning. Lower it to 0.0 for maximum consistency.

How to Experiment

1. Start with everything on — the most informed predictions
2. If predictions seem erratic, try disabling the uncertainty sections
3. If you want to test a theory, compare accuracy with and without a specific section
4. Changes take effect on the next prediction run — no restart needed
5. Disabling sections reduces token usage if you're monitoring costs

---

The Tiered Model Architecture

FIN uses a tiered system to produce the best predictions available. Each tier is a different method, and the system automatically picks the best one.

TIER 1Your Personal Model

5 quantile GBM models (q10/q25/q50/q75/q90) trained specifically on your ticker. Full prediction distribution with confidence bands. Best accuracy.

TIER 2Pooled Model

Trained across all users on the same instance. Produces median (q50) predictions only. Activates when your personal data is too sparse.

TIER 3Sector Transfer

Uses a model trained on a similar company in the same sector. q50 only, with 30% confidence penalty. Fallback when no same-ticker model exists.

TIER 4Thesis Heuristic

Always available — no training needed. Uses the AI's qualitative thesis (direction + conviction) to produce rough price estimates.

How Tiers Are Promoted

After each retrain, the system evaluates which tiers are available:

1. Tier 1: Your personal GBM models are trained for the horizon — activates when enough resolved snapshots exist
2. Tier 2: A pooled model exists — activates when user-specific training produces fewer than 3 models
3. Tier 3: A same-sector model exists — activates automatically if another stock in your sector has a trained model
4. Tier 4: Always available — the fallback when no trained model exists

Performance Gating

Even when a tier is trained, it must prove it's still accurate. The system checks the last 20 validation results:

Condition	Action
Coverage < 40% (actual price falls outside bands too often)	Tier downgraded one level
Normalized Winkler > 1.5× the bucket's own baseline (≥20%, or >50% for new buckets without enough history) — intervals too wide or inaccurate vs that instrument's norm	Tier downgraded one level

For example, if your Tier 1 model's coverage drops below 40%, the system automatically falls back to Tier 2. This prevents stale or inaccurate models from producing bad predictions.

These four thresholds — the coverage threshold, the Winkler degradation factor, the Winkler default threshold, and the Winkler floor — are per-ticker and adjustable from the dashboard under Config → ML Accuracy Gate. The values you set there are exactly what the gate reads for that (user, ticker) bucket; if you leave them unset, the system defaults (coverage 0.40, degradation 1.5×, default 50%, floor 20%) apply.

What This Means for You

• You don't need to manage tiers — the system handles promotion and demotion automatically
• Tier 1 is the goal — it gives the best predictions with full confidence bands
• Backfill accelerates Tier 1 — typically activates on the first retrain, not weeks later
• Performance gating protects you — bad models are automatically demoted

---

Other Settings Worth Knowing

Market Context (Config → Prediction Prompt)

When enabled, FIN compares your stock to market benchmarks (SPY, QQQ) and your sector ETF. It computes:

• Beta — how much your stock moves relative to the market
• Correlation — how closely your stock tracks the market
• Relative Strength — whether your stock is outperforming or underperforming

Toggle off via Config → Prediction Prompt → Market Context if you want faster processing for niche stocks that don't correlate with benchmarks.

Competitors (Competitors Page)

Add peer tickers (e.g., ORCL for SAP) for relative performance analysis:

• System auto-generates RSS feeds for competitors
• Competitor news goes through the same sentiment/surprise pipeline
• Enables the AI to distinguish sector-wide moves from company-specific signals
• Adds competitor-aware ML features (cross-impact, relative strength)

---

Tips for Getting Started

🚀

Start with defaults. They work well for most stocks. The system is designed to learn and improve on its own.

📊

Check the Jobs page to monitor what stage the pipeline is in. You can see feeds being processed, predictions running, and models training.

📈

Watch accuracy improve over time on the Predictions page. The system needs 1–2 weeks to build enough data for Tier 1 models.

📡

Tune RSS frequency based on your stock's news volume. Active stocks benefit from every 1–2 hour checks. Quiet stocks are fine at 4–6 hours.

🔬

Use Config → Prediction Prompt to experiment. Toggle sections on/off and compare accuracy over the next few days to see what data helps most for your stock.

⏱️

After 1–2 weeks, the system should have enough data for Tier 1 training. Check the Accuracy tab on the Predictions page to see how it's progressing.

---

Technical Details

Click to expand — deep-dive into algorithms, formulas, and internals

Feature Snapshot Lifecycle

1. Creation: During each prediction snapshot save is done, once per offset per horizon group. Features are assembled from multiple sources — technical and financial from the stocks information, psychology from the calculated psycology scores, calendar from the calendar events, market context from benchmark metrics, options from the options chain.

2. Resolution: After the target date passes, the accuracy tracker resolves the snapshot with actual values and calculates error percentage - llm_error_pct = ((actual − predicted) / actual) × 100`.

3. Training: Resolved snapshots are queried by with a minimum sample threshold. Features are extracted into a numeric matrix, targets into a vector, and fed to GBM training.

Backfill Algorithm

Backfill feature snapshots algorithm reconstructs feature vectors for every past trading day present that was previously stored:

• Technical features (rsi, atr, sma_crossover, price_change_5d/20d, volatility_20d): derived from stored OHLCV close series using rolling arithmetic — no re-fetching is required
• Fundamental features (pe ratio, debt to equity, revenue growth yoy, eps growth yoy, earnings_quality, free cashflow) are computed on-the-fly
• Calendar features (days to earnings, days to dividend) are derived from the stored calendar events
• Non-reconstructable features (sentiment, psychology, surprise, thesis, options, market context): set to NULL — GBM handles missing values natively via its built-in missing-value split logic
• Outcome resolution: Close price at target date is looked up from the stored data with ±5 day fallback for holidays; actual value, actual return pct, and resolved at are set immediately

The backfill triggers when len(train rows) < HORIZON_MIN_SAMPLES[h] // 2:

• Short: < 25 rows (50 // 2)
• Medium: < 20 rows (40 // 2)
• Long: < 15 rows (30 // 2)

Tier Gating Formula

If mean_coverage < 0.4 OR mean_winkler > 10 (requires ≥5 rows), the tier is downgraded:

• Tier 1 → Tier 2
• Tier 2 → Tier 3
• Tier 3 → Tier 4

Winkler Score

The Winkler (or interval score) measures both coverage and sharpness:

$$W = \begin{cases} width & \text{if } P_{low} \leq P_{actual} \leq P_{high} \\ width + \frac{2}{\alpha}(P_{low} - P_{actual}) & \text{if } P_{actual} < P_{low} \\ width + \frac{2}{\alpha}(P_{actual} - P_{high}) & \text{if } P_{actual} > P_{high} \end{cases}$$

Where width = P_high − P_low and α = 0.20 (for 80% coverage target). Lower is better — a perfect prediction scores its interval width; miss predictions are penalized proportionally to how far outside the band the actual falls.

Quantile GBM Training

5 GBM models are trained per ticker/horizon with the following settings:

• Target: What the model tries to predict - actual_return_pct (the realized return from prediction date to target date)
• Quantiles: Which percentiles are predicted - q10, q25, q50, q75, q90
• Objective: How the model learns - Quantile regression (pinball loss) is used
• Hyperparameters: The parameters that control the model behaviour - num_leaves=31, learning_rate=0.05, feature_fraction=0.8
• Validation: How we test if the model is good - Time-series cross-validation (5 folds, early stopping at 50 rounds)

The q50 gives the central prediction; the q10–q90 gap forms the confidence band.

Conformal Prediction Calibration

At inference the q10/q90 interval width is calibrated on the last 50 validation predictions to achieve ~80% coverage:

$$shortfall_i = \max(0, Q_{10,i} - y_i) + \max(0, y_i - Q_{90,i})$$

$$\alpha = \text{Quantile}(shortfall,\ 1 - coverage_{target} + coverage_{actual})$$

$$Q_{10}^{adj} = Q_{10} - \alpha, \quad Q_{90}^{adj} = Q_{90} + \alpha$$

When empirical coverage already meets or exceeds 80%, no adjustment is applied (returns 0.0).

Sentiment Decay Formula

$$w_d = e^{-\lambda \cdot \text{age\_days}}$$

Where λ is the decay rate per horizon:

• Short: λ = 0.3 (half-life ≈ 2.3 days)
• Medium: λ = 0.1 (half-life ≈ 6.9 days)
• Long: λ = 0.05 (half-life ≈ 13.9 days)

The effective weight combines decay with freshness penalty, buzz multiplier, and dimension importance:

$$w_{eff} = w \cdot (1 - \frac{j}{100}) \cdot w_f \cdot a \cdot w_b$$

Surprise Detection Pipeline

Stage 1 — Regex Pre-filter: 25 compiled regex patterns scan article text for surprise signals (earnings beats/misses, economic surprises, guidance changes, etc.). Only articles matching at least one pattern proceed.

Stage 2 — LLM Evaluation: Matched articles are evaluated on:

• is genuine — True if real surprise, not speculation
• magnitude — 0–100 (trivial to extreme)
• direction — 0–100 (very negative to very positive)
• ticker relevance — 0–100 (unrelated to directly impactful)
• market relevance — 0–100 (narrow to broad market significance)
• confidence — 0–100 (LLM certainty in its assessment)

Post-LLM filtering applies min magnitude threshold (default 20) and min ticker relevance (default 15).

Bias Weighting by Horizon

Each of the 12 psychological biases has a horizon-specific weight that controls its influence:

Bias	Short	Medium	Long	Rationale
FOMO	1.0	0.3	0.2	Short-term noise
Herding	1.0	0.3	0.2	Momentum-driven short term
Recency	1.0	0.3	0.2	Recent events dominate near-term
Availability	1.0	0.3	0.3	Easily recalled events skew short term
Loss Aversion	0.3	0.5	0.5	Stable across horizons
Confirmation	0.3	1.0	0.3	Peaks at medium term
Overconfidence	0.3	1.0	0.3	Most dangerous at medium horizon
Disposition	0.2	1.0	0.3	Medium-term holding bias
Anchoring	0.2	0.5	1.0	Reference prices dominate long term
Narrative	0.2	0.5	1.0	Stories compound over time
Ambiguity Aversion	0.2	0.3	1.0	Long-term uncertainty magnifies aversion
Regret Minimization	0.2	0.3	1.0	Long-horizon regret-avoidance grows

Tier 3 Sector-Transfer Fallback

Tier 3 has a three-level internal fallback:

1. Models from other tickers in the same GICS sector
2. A pooled model trained for this specific ticker
3. Any pooled model across all tickers

Only a single q50 model is available at Tier 3. Remaining quantile bands (q10/q25/q75/q90) are estimated from volatility. Tier 3 predictions receive a 30% confidence penalty compared to Tier 1–2.

Operational Confidence Penalties

Condition	Penalty
Stale OHLCV data	−0.2
No calendar events	−0.1
No buzz/engagement data	−0.05
No news signals (no ticker-relevant articles)	−0.2
No psychology signals	−0.05
MAPE > 5% (≥10 samples)	−0.1
Coverage < 50% (≥10 samples)	−0.1
Direction agreement (GBM = thesis)	×1.05 (boost)
Direction disagreement (GBM ≠ thesis)	×0.85
Tier 4 (heuristic-only)	×0.5

Uncertainty modulation: conf *= (1 − avg uncertainty / 100). Final confidence is clipped to [0, 1].