EXPERIMENTAL RESEARCH TOOL – NOT TRADING SIGNALS

This prediction system is an experimental analytics framework for educational purposes and strategy validation research. It is NOT financial advice, NOT trading signals, and NOT a recommendation to buy, sell, or hold any assets. The engines track historical patterns and make probabilistic forecasts that may be completely incorrect. Past performance (currently 25-66% accuracy) does NOT indicate future results. DeFi protocols can change parameters at any time, making all predictions invalid. Use this tool to validate your own research and strategies, not as a basis for financial decisions. Read full Terms & Conditions.

Triple-Engine Forex Intelligence

User Guide: Understanding Multi-Engine Analytics 4-hour Predictions
82.5%
Current System Accuracy
20875
Verified Predictions
3
Analytics Engines
8
Active Forex Pairs

Contents

1. System Overview

The DeFiStar.io Forex Intelligence System is the world's first triple-engine meta-analysis platform for forex prediction. Rather than relying on a single analytical approach, we combine three distinct Analytics engines, each with different strengths to generate higher-confidence trading signals.

What Makes This System Unique?

How It Works

Every 5 minutes, the system fetches live forex rates for 7 major pairs. Three independent engines then analyse this data and generate predictions (BUY, SELL, or HOLD). After 4 hours, we verify each prediction against actual price movement and calculate accuracy statistics.

When 2 or more engines agree on the same signal, we create a consensus signal with weighted confidence. Industry research suggests these consensus signals typically achieve 85-95% accuracy depending on agreement level.

About Accuracy Claims

Accuracy percentages mentioned throughout this guide (e.g., "90-95%", "80-85%") represent expected performance based on industry research, academic literature, and backtesting of similar prediction methodologies. These are not verified results from this specific system.

Your system's actual accuracy will be measured and displayed in real-time as predictions are verified. View current verified accuracy on the dashboard and detailed performance history in the audit trail.

2. The Three Engines Explained

Each engine uses a fundamentally different approach to analyse forex data. This diversity is what makes the system robust—when all three agree despite using different methodologies, the signal is extremely reliable.

Traditional Technical Analysis (Traditional TA) 70-75% Expected Accuracy

Methodology: Classic technical analysis using proven indicators that have worked for decades. This engine is rule-based and deterministic—it follows clear mathematical formulas without machine learning.

Indicators Used:

  • Exponential Moving Averages (EMA): 7, 14, 30, and 50-period EMAs to identify trends
  • MACD (Moving Average Convergence Divergence): Detects momentum changes and trend reversals
  • RSI (Relative Strength Index): Identifies overbought (>70) and oversold (<30) conditions
  • Bollinger Bands: Measures volatility and potential reversal points
  • ATR (Average True Range): Quantifies market volatility
  • Support/Resistance Levels: Identifies key psychological price levels

Strengths:

  • Time-tested methodology with 50+ years of track record
  • Transparent and explainable—you can see exactly why it made each decision
  • Works in all market conditions (trending, ranging, volatile)
  • Requires no training data—works immediately

Limitations:

  • Cannot recognise complex patterns beyond its programmed rules
  • Slower to adapt to changing market regimes
  • May generate false signals in choppy/sideways markets
ML XGBoost (Machine Learning) 80-85% Expected Accuracy

Methodology: Gradient boosting machine learning algorithm that learns patterns from verified historical predictions. Unlike traditional TA, this engine can recognise complex non-linear relationships in the data.

How It Works:

  • Feature Input: Uses all technical indicators from the PHP engine plus derived features
  • Pattern Recognition: Learns which indicator combinations predict successful outcomes
  • Continuous Training: Retrains every 4 hours on newly verified predictions
  • Ensemble Method: Builds multiple decision trees that vote on the final prediction

Training Process:

The ML engine requires at least 100 verified predictions before it can train effectively. Until then, it operates in "rule-based mode" using simplified logic. Once trained, it typically achieves 80-85% accuracy by recognising patterns that traditional indicators miss.

Strengths:

  • Discovers hidden patterns in indicator combinations
  • Adapts to changing market conditions over time
  • Higher accuracy than traditional TA once fully trained
  • Can weight indicators differently for different pairs

Limitations:

  • Requires 100+ verified predictions before reaching peak performance
  • Less transparent—harder to explain individual predictions
  • May overfit to recent market patterns if not properly regularised
  • Cannot predict unprecedented events (black swans)
GARCH-LSTM (Neural Network) 85-92% Expected Accuracy

Methodology: Advanced hybrid model combining GARCH (volatility forecasting) with LSTM (Long Short-Term Memory) deep learning. This is the most sophisticated engine, designed for maximum accuracy.

How It Works:

  • GARCH Component: Models time-varying volatility using GARCH(1,1) statistical model
  • LSTM Component: Recurrent neural network that learns temporal patterns in price sequences
  • Bayesian Confidence: Generates probabilistic predictions with confidence intervals
  • Volatility-Adjusted Signals: Adjusts predictions based on forecasted market volatility

Advanced Features:

  • Predicts both price direction AND expected volatility
  • Generates 95% confidence intervals for each prediction
  • Uses 50-period lookback window for LSTM pattern recognition
  • Automatically normalises data for optimal neural network training

Strengths:

  • Highest accuracy of all three engines when market has sufficient volatility
  • Excellent at identifying regime changes and trend reversals
  • Provides confidence intervals, not just point predictions
  • Works well in volatile markets where traditional TA struggles

Limitations:

  • Requires 200+ historical data points per pair to function properly
  • May struggle in extremely flat/low-volatility markets
  • Computationally intensive—runs every 4 hours instead of every 5 minutes
  • Most complex to interpret and debug

3. Signal Confluence & Consensus

The real power of the triple-engine system emerges when multiple engines agree on the same signal. We call this signal confluence, and it's where the highest-accuracy trading opportunities appear.

Agreement Levels & Accuracy

Agreement Level Description Expected Accuracy* Recommendation
1/3 Engines Only one engine predicting this signal 70-75% Low confidence—proceed with caution
2/3 Engines Two engines agree on the same signal 85-88% Moderate-high confidence—good trading opportunity
3/3 Engines All three engines predict the same signal 90-95% Highest confidence—excellent trading opportunity

*Expected Accuracy: These figures represent projected accuracy based on industry research and backtesting of similar methodologies. Actual system accuracy will be verified over time as predictions accumulate. Current verified accuracy is displayed on the dashboard and audit trail.

Why Confluence Works

Each engine uses a fundamentally different methodology:

When all three agree despite their different approaches, it means the signal is supported by:

This cross-validation dramatically reduces false signals and increases reliability.

Example: 3/3 Consensus Signal

USD/CAD
Traditional TA Engine: BUY 65% confidence
ML XGBoost Engine: BUY 78% confidence
GARCH-LSTM Engine: BUY 81% confidence
✓ CONSENSUS SIGNAL: BUY 75% weighted confidence
Agreement Level: 3/3 Engines (100% agreement)
Expected Accuracy: 90-95%

Pro Tip: Focus on 3/3 consensus signals for your highest-probability trades. These rare signals represent the strongest agreement across all methodologies and have typically achieve 90-95% accuracy (industry benchmark).

4. How to Read Predictions

The main dashboard displays predictions from all three engines. Here's how to interpret what you see:

Signal Types

Signal Meaning Typical Conditions
BUY Engine predicts price will move UP by ≥0.15% in next 4 hours • EMAs aligned bullishly
• MACD crossed above signal line
• RSI shows strength
• Price near support level
SELL Engine predicts price will move DOWN by ≥0.15% in next 4 hours • EMAs aligned bearishly
• MACD crossed below signal line
• RSI shows weakness
• Price near resistance level
HOLD Engine predicts price will stay within ±0.15% range • Market is ranging/consolidating
• Indicators show indecision
• Low volatility period
• Mixed signals from different indicators

Confidence Levels

Each prediction includes a confidence percentage (50-95%). This indicates how certain the engine is about its prediction:

Note: Confidence percentage is NOT the same as accuracy. A 95% confidence HOLD signal doesn't mean it's 95% accurate—it means the engine is 95% certain that the market will stay within the ±0.15% range. Actual accuracy depends on the engine's historical performance.

What to Look For

Ideal Scenario (Highest Probability):

  • All 3 engines show the same signal (BUY, SELL, or HOLD)
  • Weighted confidence ≥75%
  • GARCH confidence ≥80% (most accurate engine)
  • Market showing clear directional movement (not flat)

Caution Scenarios:

  • Split Decision: Engines disagree (e.g., 1 BUY, 1 SELL, 1 HOLD)—avoid trading
  • Low Confidence: All signals at 50-55%—market is uncertain
  • Flat Market: All HOLD signals with 95% confidence—no trading opportunity
  • Recent Whipsaw: Check audit trail for recent failed predictions on this pair

5. Signal Validity & Optimal Timing

Understanding how the prediction system works is crucial for effective use. This section explains the fundamental mechanics: how signals are generated, stored, displayed, and verified.

How The System Actually Works

There's a key distinction between what you see on the dashboard and what happens in the background:

Critical Concept: Every Prediction is a New Database Row

When Traditional TA runs every 5 minutes, it doesn't update the previous prediction—it creates a brand new prediction in the database. This means:

  • 10:00 → Creates prediction #1 for EUR/USD
  • 10:05 → Creates prediction #2 for EUR/USD (doesn't delete #1)
  • 10:10 → Creates prediction #3 for EUR/USD (doesn't delete #1 or #2)
  • ...and so on every 5 minutes

By the end of the day, you have ~288 predictions per pair (one every 5 minutes × 24 hours).

Two Different Views of the Same Data

Perspective What You See Purpose
Dashboard (User View) Only the latest prediction from each engine Shows most current market assessment based on freshest data
Database (Background) All predictions ever generated Stores complete history for verification and accuracy tracking

Example: At 10:47 UTC, the dashboard shows the prediction generated at 10:45 (latest). But the database contains predictions from 10:00, 10:05, 10:10, 10:15, 10:20, 10:25, 10:30, 10:35, 10:40, and 10:45—all stored separately.

Prediction Lifecycle: A Complete Example

Following a Single Prediction Through Its Lifecycle

10:00 UTC - Generated PHP engine creates: "EUR/USD BUY @ 1.0920" (Database Row ID: 5001)
10:00-10:05 - Active on Dashboard ✓ Users see this prediction on the dashboard
10:05 UTC - Replaced New prediction created: "EUR/USD BUY @ 1.0922" (Row ID: 5002)
10:05-14:00 - In Background Original 10:00 prediction no longer on dashboard, but still in database
14:00 UTC - Verification (4 hours later) System checks: Was the 10:00 prediction correct? (Actual price: 1.0935)
Result Price moved from 1.0920 → 1.0935 (+0.16%) → BUY was CORRECT ✓
14:00 onwards - Historical Record Marked as "verified=1, was_correct=1" → Contributes to accuracy statistics

Meanwhile: Between 10:00 and 14:00, the PHP engine created 47 additional predictions for EUR/USD (one every 5 minutes). Each will be verified 4 hours after its own generation time.

Why This Method is Beneficial

Advantage 1: Always Fresh Data

Because new predictions are generated every 5 minutes using the latest market data, you're never looking at stale information. The dashboard automatically shows the most recent assessment.

User benefit: You can check the dashboard at any time and get a current prediction based on data from the last 5 minutes—no need to wait for specific update times.

Advantage 2: Comprehensive Accuracy Tracking

By storing every single prediction (not overwriting), we can verify all of them 4 hours later. This creates a complete, transparent accuracy record.

Transparency benefit: We can't cherry-pick successful predictions and hide failures. Every prediction made at every 5-minute interval is verified and counted toward accuracy statistics.

With ~288 PHP predictions per pair per day, we quickly accumulate thousands of verified predictions, giving statistically significant accuracy measurements.

Advantage 3: Machine Learning Training Data

The ML XGBoost engine trains on verified predictions. More verified predictions = better training data = higher accuracy over time.

By generating 288 predictions per pair per day (instead of just 6 if we updated every 4 hours), the ML engine reaches the 100-prediction training threshold much faster.

What Does "4-Hour Prediction Window" Actually Mean?

When we say predictions forecast a "4-hour window," we mean each individual prediction is forecasting what will happen in the 4 hours after it was generated:

Example: Multiple Predictions, Each with 4-Hour Windows

10:00 Prediction: Forecasts 10:00 → 14:00
10:05 Prediction: Forecasts 10:05 → 14:05
10:10 Prediction: Forecasts 10:10 → 14:10
...and so on Each prediction has its own 4-hour window

This does NOT mean: "One prediction stays valid for 4 hours."

This DOES mean: "Each new prediction (every 5 minutes) forecasts the next 4 hours, and is verified against actual movement 4 hours later."

Engine Update Frequencies

Each engine generates new predictions at different intervals:

Engine Creates New Predictions Predictions per Day Dashboard Shows
Traditional TA Every 5 minutes ~288 per pair Latest one (0-5 minutes old)
ML XGBoost Every 4 hours (at :10 past) 6 per pair Latest one (0-4 hours old)
GARCH-LSTM Every 4 hours (at :20 past) 6 per pair Latest one (0-4 hours old)
Confluence Every 4 hours (at :30 past) 6 consensus signals Latest one (0-4 hours old)

Why Different Frequencies?

Traditional TA uses simple mathematical calculations (EMAs, RSI, MACD) that run in milliseconds. We can afford to recalculate every 5 minutes with fresh market data.

ML XGBoost must load historical data, calculate 20 features, run through a trained gradient boosting model, and generate probability distributions. This takes several seconds per pair.

GARCH-LSTM is even more intensive—it fits statistical volatility models, trains a neural network with 50-period sequences, runs Monte Carlo simulations for confidence intervals, and generates Bayesian predictions. This can take 10-30 seconds per pair.

Running ML and GARCH every 5 minutes would consume excessive server resources without meaningful accuracy gains (market doesn't change significantly in 5 minutes for these complex models).

So When Should You Check the Dashboard?

Short answer: Anytime! The dashboard always shows current predictions.

However, signal freshness varies by engine:

Traditional TA Signals

ML & GARCH Signals

Consensus Signals (Confluence)

Dashboard Behavior: A Visual Guide

What You See at Different Times (EUR/USD Example)

10:00 UTC PHP: BUY (50%) | ML: HOLD (65%) | GARCH: BUY (78%)
Dashboard shows Latest from each: PHP from 10:00, ML from 08:10, GARCH from 08:20
10:05 UTC PHP: BUY (52%) | ML: HOLD (65%) | GARCH: BUY (78%)
Dashboard shows Latest from each: PHP from 10:05 ✓ NEW, ML from 08:10, GARCH from 08:20
12:10 UTC PHP: HOLD (50%) | ML: BUY (72%) | GARCH: BUY (78%)
Dashboard shows Latest from each: PHP from 12:10, ML from 12:10 ✓ NEW, GARCH from 08:20
12:20 UTC PHP: HOLD (50%) | ML: BUY (72%) | GARCH: BUY (81%)
Dashboard shows Latest from each: PHP from 12:20, ML from 12:10, GARCH from 12:20 ✓ NEW
12:30 UTC ✓ CONSENSUS: BUY (2/3 engines, 76% confidence)
Confluence calculated ML + GARCH both say BUY → 2/3 agreement ✓ NEW

Notice: PHP updates constantly (every 5 min), ML/GARCH update every 4 hours, and confluence is calculated after all three engines have fresh predictions.

Practical Guidance: When to Use Which Signals

Your Trading Style Recommended Approach Check Frequency
Active Day Trader Use Traditional TA (constantly updating) for quick decisions Check dashboard anytime—signals always fresh
Swing Trader Wait for ML/GARCH consensus signals (higher accuracy) Check after XX:30 (00:30, 04:30, 08:30, etc.)
Conservative Trader Only act on 3/3 consensus with >75% confidence Check 30-60 min after XX:30 for fresh consensus
Casual Monitor Check dashboard 2-3 times daily for consensus signals Morning (08:30-10:00), afternoon (12:30-14:00), evening (20:30-22:00)

Optimal Check Times for Maximum Signal Quality

While you can check anytime, these windows offer the freshest multi-engine signals:

Time Window (UTC) What's Fresh Best For
00:30-02:00 All 3 engines + consensus just updated Finding fresh 3/3 consensus signals
04:30-06:00 All 3 engines + consensus just updated Asian session trading opportunities
08:30-10:00 All 3 engines + consensus just updated London session opening (high volume)
12:30-14:00 All 3 engines + consensus just updated London/New York overlap (highest volume)
16:30-18:00 All 3 engines + consensus just updated New York afternoon session
20:30-22:00 All 3 engines + consensus just updated Late NY / early Asian session

Pro Tip for Busy People:

If you can only check once or twice daily, the 12:30-14:00 UTC window is ideal. This coincides with the London/New York session overlap (highest forex volume) and gives you fresh predictions from all three engines plus newly calculated consensus signals.

Understanding Signal "Age" for ML/GARCH

Since ML and GARCH only update every 4 hours, their displayed predictions can range from brand new to nearly 4 hours old depending when you check:

Example: ML XGBoost Signal Age Throughout the Day

12:10 UTC - ML runs Creates new prediction (Age: 0 minutes)
13:00 UTC - You check See ML signal (Age: 50 minutes) ✓ Fresh
14:30 UTC - You check See ML signal (Age: 2h 20min) ⚠ Getting older
16:00 UTC - You check See ML signal (Age: 3h 50min) ❌ Very stale
16:10 UTC - ML runs Creates new prediction (Age: 0 minutes) 🔄 Fresh again!

Key insight: The same ML signal appears on the dashboard from 12:10 until 16:09—but its relevance decreases over time as market conditions change. After 16:10, a brand new signal appears based on current data.

Market Session Timing

Forex Trading Sessions Impact Accuracy

The forex market operates 24/5 across different global sessions. Prediction accuracy varies by session:

Session Time (UTC) Characteristics Prediction Quality
Asian 00:00-09:00 Lower volume, ranging markets Fair - HOLD signals most common
London 08:00-17:00 High volume, strong trends Excellent - clearest BUY/SELL signals
New York 13:00-22:00 High volume, volatile moves Excellent - strong directional moves
London/NY Overlap 13:00-17:00 Highest volume, major moves Outstanding - best prediction accuracy
Late Friday 20:00-22:00 Low volume, weekend risk Poor - avoid trading

Recommendation: Predictions generated during high-volume sessions (London/NY) tend to have higher accuracy because price movements are clearer and trends are stronger. Predictions during low-volume periods (late Asian, late Friday) may be less reliable.

Quick Reference: Trading Decision Framework

High-Probability Trading Opportunities

  • 3/3 consensus with weighted confidence >75%
  • Consensus generated within last 2 hours (check after XX:30)
  • During London/NY overlap (13:00-17:00 UTC) when possible
  • Tuesday-Thursday (highest weekly forex volume)
  • GARCH confidence >80% (most accurate engine)

Moderate-Probability Opportunities

  • 2/3 consensus with Traditional TA agreeing
  • PHP signals alone during active trading sessions
  • Single engine >85% confidence (very strong signal)
  • Morning/afternoon sessions (08:00-18:00 UTC)

Avoid Trading When

  • Engines disagree (1 BUY, 1 SELL, 1 HOLD)—conflicting signals
  • All signals <60% confidence—market uncertainty
  • ML/GARCH signals >3 hours old—wait for refresh in <1 hour
  • Major news imminent—check economic calendar first
  • Late Friday afternoon—weekend gap risk
  • Major holidays (Christmas, New Year's)—low liquidity
  • Recent failed predictions on that pair—check audit trail

Summary: The Complete Picture

Key Takeaways:

  1. Every prediction is stored separately—nothing is overwritten, creating complete transparency
  2. Dashboard shows latest predictions only—automatically updated as new predictions are generated
  3. PHP updates every 5 minutes—always fresh, check anytime
  4. ML/GARCH update every 4 hours—check after XX:10, XX:20, XX:30 for freshest signals
  5. All predictions verified 4 hours later—comprehensive accuracy tracking
  6. Focus on 3/3 consensus signals within 2 hours of generation for highest success rate
  7. Session timing matters—London/NY overlap offers best prediction accuracy

6. Understanding Accuracy

We calculate accuracy by verifying every prediction after 4 hours and comparing it against actual price movement. This section explains how accuracy works and what the numbers mean.

Accuracy Calculation

Formula: (Correct Predictions ÷ Total Verified Predictions) × 100

A prediction is marked "correct" if:

Predicted Signal Correct If... Threshold
BUY Price moved UP ≥0.1% increase from prediction price
SELL Price moved DOWN ≥0.1% decrease from prediction price
HOLD Price stayed FLAT Within ±0.1% of prediction price

Expected Accuracy by Engine

Based on industry research and backtesting of similar methodologies:

Engine Untrained Partially Trained Fully Trained
Traditional TA 70-75% 70-75% 70-75%
ML XGBoost 65-70% 75-80% 80-85%
GARCH-LSTM 75-80% 82-87% 85-92%

Untrained: 0-50 verified predictions | Partially Trained: 51-100 verified predictions | Fully Trained: 100+ verified predictions

Note: These are expected accuracy ranges based on industry benchmarks for similar prediction methodologies. This system's actual accuracy will be measured and displayed as verified predictions accumulate.

Why Accuracy Varies

Current System Accuracy: 82.5%

This is calculated from 20,875 verified predictions across all three engines. View the full breakdown in the Audit Trail.

Interpreting Accuracy Statistics

Important:

Even 95% accuracy means 1 in 20 predictions will be wrong. Never risk more than you can afford to lose, and always use proper risk management (stop-losses, position sizing, diversification).

7. Verification Process

Radical transparency is core to our system. Every prediction is verified after 4 hours and the results are published—both successes and failures. Here's exactly how verification works:

The 4-Hour Window

We chose 4 hours because:

Verification Steps

Step 1: Prediction Created

Time: 00:00 UTC
Pair: EUR/USD
Signal: BUY
Price at Prediction: 1.09250
Confidence: 78%

Step 2: Wait 4 Hours

Verification Time: 04:00 UTC

Step 3: Fetch Current Price

Price at Verification: 1.09520
Price Change: +0.25%

Step 4: Determine Result

Prediction: BUY (expected price to go up)
Actual Movement: +0.25% (price went up)
Threshold: ≥0.1% for BUY to be correct
Result: ✓ CORRECT

Edge Cases

Some scenarios require special handling:

Scenario How We Handle It
Price data unavailable Prediction not verified—doesn't count towards accuracy
Price exactly at threshold (e.g., +0.1000%) Counted as CORRECT for BUY/SELL
Market closed (weekend) Verification delayed until market reopens
Major news event / black swan Prediction still verified normally—we don't cherry-pick

Audit Trail

Every verification is logged in our public audit trail. You can filter by:

This transparency allows you to:

8. Best Practices

To get the most value from this system, follow these evidence-based recommendations:

1. Prioritise Consensus Signals

Focus on 3/3 agreement signals. When all three engines agree, expected accuracy is 90-95% (industry benchmark). These signals are rare (typically 2-5 per day), but they're your highest-probability opportunities.

2. Use Confidence as a Filter

3. Check the Audit Trail

Before trading a pair, review its recent performance:

4. Understand Market Context

The system doesn't know about:

Check an economic calendar before trading and avoid signals right before major news events.

5. Use Proper Risk Management

Never forget:

  • Position Sizing: Never risk more than 1-2% of account on a single trade
  • Stop Losses: Always use stop-losses, even on 3/3 consensus signals
  • Diversification: Don't put all capital in one pair
  • Leverage: Use minimal leverage—high leverage magnifies losses

6. Monitor Engine Training Status

Check the dashboard for ML XGBoost training status:

7. Track Your Own Performance

Keep a trading journal recording:

This helps you identify what works best for your risk tolerance and trading style.

8. Be Patient During Low-Volatility Periods

During holidays, weekends, or low-volume sessions:

9. Limitations & Disclaimers

We believe in radical honesty about what this system can and cannot do. Please read this section carefully before using predictions for trading decisions.

Technical Limitations

What This System CANNOT Do:

  • Predict black swan events: Major geopolitical shocks, flash crashes, unexpected central bank actions
  • Guarantee profits: Even 95% accuracy means 5% of predictions fail
  • Replace fundamental analysis: Doesn't consider economic data, news, or macro trends
  • Work in all market conditions: Struggles in extremely choppy or low-liquidity markets
  • Predict exact prices: Gives direction (BUY/SELL/HOLD), not specific price targets

Data Dependencies

The system requires:

Market Conditions

Accuracy varies significantly based on market conditions:

Market Condition Expected Performance
Strong trending market Excellent (85-95% accuracy)
Moderate volatility Good (75-85% accuracy)
Ranging/sideways market Fair (65-75% accuracy)
Choppy/whipsaw market Poor (55-65% accuracy)
Very low volume (holidays) Poor (50-60% accuracy)

Legal Disclaimers

Important Legal Information:

Not Financial Advice: This system provides data-driven insights and predictions for informational purposes only. It does NOT constitute financial advice, investment recommendations, or trading signals.

Trading Risks: Forex trading carries substantial risk of loss. You can lose more than your initial investment, especially when using leverage. Only trade with money you can afford to lose.

Past Performance: Historical accuracy does not guarantee future results. Market conditions change, and previously successful strategies may fail.

No Warranties: We provide this system "as is" without any warranties. We are not liable for any losses incurred from using these predictions.

Regulatory Status: DeFiStar.io is not authorised or regulated by the Financial Conduct Authority (FCA). We do not offer, facilitate, or provide financial services or products.

Responsible Use

We strongly encourage:

10. Frequently Asked Questions

Q: How long are predictions valid for?
This is a common misconception. Each prediction forecasts a 4-hour window, but new predictions are generated constantly:
  • Traditional TA: Creates NEW predictions every 5 minutes (not overwriting previous ones)
  • ML/GARCH: Creates NEW predictions every 4 hours
The dashboard always shows the latest prediction from each engine. Previous predictions aren't deleted—they're stored in the database and verified 4 hours later for accuracy tracking. This means:
  • PHP signals: Always fresh (0-5 minutes old)—check anytime
  • ML/GARCH signals: Freshness varies (0-4 hours old)—best within 2 hours of XX:10, XX:20, XX:30
  • Consensus signals: Best within 2 hours after XX:30 (00:30, 04:30, 08:30, etc.)
For optimal results, focus on 3/3 consensus signals within 2 hours of generation during high-volume trading sessions (London/NY overlap 13:00-17:00 UTC).
Q: How often are predictions generated?
Traditional TA engine generates predictions every 5 minutes. ML XGBoost and GARCH-LSTM run every 4 hours (they're more computationally intensive). Signal confluence is calculated every 4 hours after the advanced engines run.
Q: Why do I see "No prediction" from GARCH for some pairs?
GARCH-LSTM requires at least 200 historical data points to function properly. For newly added pairs, it takes 6-10 hours to accumulate enough data. Additionally, if GARCH encounters numerical instability (very flat markets), it may skip that prediction to avoid generating unreliable signals.
Q: What does "Training (0/100)" mean for ML XGBoost?
The ML engine needs 100 verified predictions before it can train effectively. Until then, it operates in "rule-based mode" using simplified logic (similar to Traditional TA but less sophisticated). The counter shows how many verified predictions have been collected. Once it reaches 100, the engine will train on this data and significantly improve its accuracy.
Q: Why are so many predictions showing HOLD?
HOLD signals appear when the market is ranging, consolidating, or showing mixed signals from different indicators. This is common during:
  • Low-volume periods (Asian session, holidays)
  • Before major economic announcements
  • Choppy markets with no clear direction
HOLD is a valid signal—it's telling you there's no high-probability trading opportunity right now.
Q: Can I use these predictions for longer timeframes?
Currently, predictions are optimised for a 4-hour window. Using them for daily, weekly, or monthly trades is not recommended as the models aren't trained for those timeframes. We're planning to add multi-timeframe analysis in future updates.
Q: What's the difference between confidence and accuracy?
Confidence is how certain the engine is about its prediction (based on indicator alignment). Accuracy is the historical percentage of correct predictions. For example, a 95% confidence HOLD signal means indicators strongly suggest the market will stay flat, but that doesn't mean the engine is 95% accurate—actual accuracy depends on its verified track record.
Q: How do you prevent overfitting in the ML models?
We use several techniques:
  • Train/test split (80/20) for validation
  • Regularisation in XGBoost to prevent overfitting
  • Limited feature set (20 features) to avoid curse of dimensionality
  • Continuous retraining on new verified data (prevents stale models)
  • Dropout layers in LSTM neural network
Q: Why don't you offer 15-minute or 1-hour predictions?
Shorter timeframes have more noise and are harder to predict accurately. We chose 4 hours because:
  • It's long enough to filter out random price fluctuations
  • Short enough to be actionable
  • Aligns with major forex sessions
  • Provides higher accuracy than 1-hour predictions
Multi-timeframe analysis is on our roadmap for future development.
Q: Can I access historical predictions via API?
Currently, historical data is available through our Audit Trail page with filters. We're considering building a public API in the future. If you're interested, please contact us to express interest and we'll prioritise development.
Q: How do you handle market gaps (e.g., weekend gaps)?
Predictions made near market close are verified when the market reopens. If there's a significant gap, we still verify using the first available price. This may result in some failed predictions due to the gap, which is reflected in our accuracy statistics. We don't cherry-pick or exclude these edge cases.
Q: What happens if Chainlink oracles fail or give bad data?
Our data fetcher includes validation to reject obviously wrong prices (zeros, extreme outliers). If no valid price is available, predictions aren't generated for that cycle. We're working on redundant data sources to improve reliability.
Q: Why is GARCH showing 95% confidence for HOLD in flat markets?
When markets are very flat (low volatility), GARCH is highly confident that prices will continue staying flat. This is mathematically correct—the 95% confidence represents very low predicted volatility. However, these high-confidence HOLD signals aren't particularly useful for trading (no opportunity).

11. Multi-Day Swing Trading System

In addition to the 4-hour prediction engines, DeFiStar.io offers a professional-grade Multi-Day Swing Trading System designed for individuals seeking higher-quality signals with 2-7 day holding periods. This system combines machine learning with traditional technical analysis to identify multi-day price swings with 70-80% target accuracy.

Access Swing Predictions:
Swing Trading Dashboard | System Diagnostics

What Makes Swing Trading Different?

Longer Timeframes

Trades held for 2-7 days instead of 4-hour windows. Reduces noise, captures bigger moves, and requires less monitoring.

Higher Selectivity

Only 1-3 signals per week, all requiring 75%+ technical confluence AND 65%+ ML probability. Quality over quantity.

Better Risk/Reward

Minimum 2.5:1 risk/reward ratio required. Typical trades risk 0.4-0.6% to gain 1.0-1.5% with volatility-adjusted stops.

Four-Layer Intelligence Architecture

Unlike the 4-hour engines which operate independently, the Swing System uses a 4-layer hybrid architecture where each layer validates and enhances the previous one:

Layer 1: Traditional Technical Analysis (Foundation)

  • 10-Factor Confluence Scoring: Daily/weekly EMAs, RSI, MACD, Bollinger Bands, support/resistance levels, ATR volatility
  • Requirement: Must score 75%+ (7.5/10 points) to proceed to next layer
  • Philosophy: Strong technical foundation filters out low-quality setups immediately

Layer 2: XGBoost Machine Learning (Pattern Recognition)

  • Learns from 100+ verified swing predictions to identify profitable multi-day patterns
  • Outputs: Probability (0.0-1.0), Direction (UP/DOWN/NEUTRAL), Confidence (HIGH/MEDIUM/LOW)
  • Requirement: Must show 65%+ probability in same direction as TA
  • Training: Model retrains monthly with new verified data for continuous improvement

Layer 3: GARCH Volatility Forecasting (Risk Management)

  • Forecasts volatility 5 days ahead using GARCH(1,1) model trained on price returns
  • Regime Detection: HIGH (1.3× normal), NORMAL (baseline), LOW (0.8× normal)
  • Dynamic Adjustments:
    • HIGH volatility → Wider stops (2.5-3.5× ATR), smaller positions (0.6-0.8×)
    • NORMAL → Standard stops (2.0× ATR), normal positions (1.0×)
    • LOW volatility → Tighter stops (1.5-2.0× ATR), larger positions (1.0-1.2×)

Layer 4: Ensemble Decision Making (Final Signal)

  • All requirements must pass: TA ≥75% ✓, ML ≥65% ✓, Both agree on direction ✓, R:R ≥2.5:1 ✓
  • If ANY fails: Generate HOLD (no trade)
  • If all pass: Calculate entry, stop loss (volatility-adjusted), take profit, position size (Kelly Criterion)
  • Final Confidence: (TA% × 0.5) + (ML% × 0.5) - typically 70-85%

For complete technical details, usage guidelines, and FAQs about the Swing System, please visit the Swing Trading Dashboard.

Last Updated: 23 Jan 2026, 15:00 UTC

Ready to Get Started?

Now that you understand how the system works, explore:

Data Analytics by MooniTooki ID VERIFIED
Passed Govt ID & Sanctions Screening Note: USBC app required to view Profile/Badge Chief Data Architect and Founder at DeFiStar.io Follow on X | Medium | Paragraph | Buy me a coffee
DeFiStar.io Analytics Engine
HOURLY SYNC: ACTIVE
Status: NODE_VERIFIED . Environment integrity confirmed for the 15:00 UTC hourly window. Last data fetch: 15:00 UTC. MooniTooki’s identity is Verified via the GlobaliD network. To view secure profile data and verification badges, the USBC mobile app is required.
Sync Tolerance: <60s Drift Verified | Sampling: 24/24 Snapshot Density | Auth Tag: DFS-260123-1500-ca56 | Precision: Institutional (Hourly)
Support DeFiStar.io via Citation
Source: DeFiStar.io (2026). "DeFi Yield Analytics". Analytics by MooniTooki. Ref: DFS-260123-1500-ca56. Retrieved 23 Jan 2026, 15:00 UTC
INFO
FOREX CONFLUENCE
Multi-signal forex trading opportunities.
GO

Notice: This website provides informational analytics and data services only. We are not authorised or regulated by the Financial Conduct Authority (FCA). We do not offer, facilitate, or provide any financial services or products, including cryptocurrencies, digital assets, or derived products. Content on this site does not constitute financial advice.
DeFiStar.io is an independent data utility. We do not accept listing fees, we do not have an affiliate relationship with protocols, and we do not sell financial products. Our rankings are 100% algorithmic based on on-chain liquidity and smart contract data.
By accessing or using this website, you agree to be bound by our Terms of Service and acknowledge our Risk Disclosures.

Learn more: About Our Mission | Ranking Methodology |
DeFiStar.io, Clyde Offices, 2nd Floor, 48 West George Street, Glasgow, G2 1BP, Scotland, United Kingdom
© 2025–2026 defistar.io