I am reverse-engineering a synthetic OTC binary options pric...

Prompt

I am reverse-engineering a synthetic OTC binary options price feed on the Quotex platform. I need help finding a 100% accurate trading edge. ============================================ PLATFORM FACTS (confirmed through testing): ============================================ 1. OTC prices are server-generated (no real buyers/sellers) 2. Prices track real Binance crypto prices with a variable lag (5-60 seconds) 3. The lag changes per session (Sydney/Tokyo/London/New York) 4. The feed is GLOBAL (same prices shown to all users worldwide) 5. ~54% of ticks are flag=0 (synthetic noise), ~46% are flag=1 (real price moves) 6. All 62 OTC assets are generated in a SINGLE batch every ~7ms (confirmed via timestamp analysis) 7. The same 15-character UP/DOWN sequences appear across 16+ different assets (shared PRNG confirmed) 8. A global 50/50 balancer maintains UP/DOWN ratio within ±0.5% across 170,000+ ticks 9. Individual assets can drift up to ±4.7% bias but the GLOBAL sum stays near zero 10. The server runs at approximately 140Hz tick generation rate 11. OTC asset prices = Binance price × fixed scale factor (confirmed per asset) 12. Scale factors are stable (e.g. DOT = 0.9245, BTC = 1.0184, ETH = 0.9973) ============================================ WHAT I TESTED BEFORE (74 tests, all documented): ============================================ - Multiple lag windows (1s, 3s, 5s, 10s, 15s, 30s, 60s) - Move size filters (small vs large Binance moves) - Move consistency filters (clean vs choppy movement) - Candle zone timing (early/mid/late in candle) - 5-minute session position filters - Tick density filters (busy vs quiet market) - Body strength filters (strong vs weak candle bodies) - Drift ratio between OTC and Binance - Multi-asset simultaneous confirmation - Volume patterns - Result chain patterns (win/loss sequences) - Spike detection with size categorization - Fake tick filtering (flag=0 vs flag=1) - Cross-asset correlation - Time-based cycles and clock rotations - PRNG cracking attempts (LCG, XOR, MT19937) — all failed - Packet layer analysis (nothing hidden there) - Global balance/governor theories - N-gram direction sequences - Consecutive run analysis - Anti-momentum trading - Candle body limit boundaries - Streak boundaries (max consecutive same direction) - Wick ratio boundaries - Session deviation limits - Tick count per candle - Time rotation explorer - Second-of-minute bias - Timestamp seed cracker - Global tick modulo patterns - Massive jump detector - Spike + candle zone combinations - Pre-spike context analyzer - Spike density analysis - Fake spike filter (500ms confirm) - Spike size ladder - Mode detection (oscillating vs trending) - Loss predictor (after N losses) - Pressure indicator (ups vs downs) ============================================ WHAT I TESTED WITH AI ASSISTANT (22+ versions): ============================================ v1-v6: Direct lag trading (Binance direction → predict OTC) Result: 71% base, degrades to 50% with more samples v7-v8: Lag with filters (F1 conflict, F2 consensus, F3 spike, F4 move size, F5 boundary) Result: 96.4% on 56 samples, dropped to 36% next day (session-specific) v9: Catch-up gap prediction (where OTC SHOULD be vs where it IS) Result: 45% — the gap is already priced in v10-v12: Direction match correlation with fresh 30s windows Result: 50-55% — correlation doesn't predict v13-v14: Event replay detection (Binance jumps → OTC echoes) Result: BTC 68-75%, ETH 55% — decent but unstable v15: Multi-angle tester (6 different signal angles in parallel) Result: B_frequency 71% on BTC, 23% on ETH (inverse) v16: Direction flip trading Result: 52% — flips don't predict v17: OTC intrinsic rhythm (flip cycle timing) Result: 37.5% — rhythm is noise v18: Candle-by-candle live vs OTC across all assets and timeframes Result: 55.3% overall, best individual combo 85.7% on 14 samples v19: Size-bucketed lag (different lags per move magnitude) Result: Very few signals, 33% v20: Balance force + flag=0 burst patterns Result: ~50% — balancer doesn't predict next tick v21: Global crowd bias across all 62 assets Result: Confirmed balancer but no trading edge v22: Universal balancer observer (all 62 assets) Result: Confirmed architecture but no edge ============================================ DEEP PROBES PERFORMED: ============================================ 10 basic probes: - Time-of-day bias → NO bias found - Trigger threshold → NO clear trigger - Correction speed → 898 ticks median recovery - Correction carrier → NO specific carrier (distributed evenly) - Batch internal structure → NO structure found - Seed rotation detection → NO clear rotation pattern - Flag=0 burst → direction correlation → STRONG on 32 samples but FAILED split validation - Price digit patterns → FAILED split validation - Batch position bias → FAILED split validation - Session regime → Stable, no shifts 6 deep combinatorial probes (pandas, 181K directional ticks): - Asset × Hour × Pattern → 0 validated edges - Asset × Second-of-Minute → 0 validated edges - Consecutive same-price + direction → 0 validated edges - 5-tick momentum × asset → 0 validated edges - Asset × Second × Pattern (triple combo) → 0 validated edges - Asset × Last-Digit × Pattern → 0 validated edges 5 rule-trigger probes: - Round-number attractor → 90%+ on next-tick but drops to 50-70% at real expiries (5-60s) - Volatility injection direction → NO bias - Trend age reversal cliff → NO cliff found - Balance emergency trigger → NO clear trigger - Payout correlation → weak negative correlation (-0.14) ============================================ WHAT PARTIALLY WORKED: ============================================ 1. SECOND-OF-MINUTE × LAG × ASSET combinations During Sydney session: BTC 30s lag at seconds 19/20 = 94% win (34/36 samples) During London session: LTC inverse at second 14 = 87.5% (different edges) BUT: edges are SESSION-SPECIFIC and change every few hours 2. Manual trading with "golden rules" (user's visual filtering): BTC: 10/10 wins, ETH: 9/12 wins during one session Overall script accuracy same session: 53.3% The human added intuitive filters the code doesn't have 3. Session high/low boundary reversal (overnight parallel_hunter): 87-100% on 17-36 samples across multiple assets But drops to 50-70% at real trading expiries (5-60s) ============================================ THE 50/50 BALANCER (mathematically confirmed): ============================================ - Across 170,000+ flag=1 ticks on 62 assets: bias = -0.212% to -0.573% - All 62 assets generated in single 1ms batches (2,592 occurrences) - Same 15-char direction sequences appear in 16 assets (shared PRNG) - Server clock: ~140Hz (7ms average gap between batches) - Per-asset bias allowed up to ±4.7% but global sum constrained near zero - No pairwise correlation between assets (each appears independent) - Recovery from imbalance: ~898 ticks median THE BALANCER IS THE WALL. Every edge we find gets balanced out. It works at the AGGREGATE level across all 62 assets, not per-asset. Individual ticks appear random even though the formula (OTC = Binance × scale + noise) holds perfectly at the description level. ============================================ THE CORE PROBLEM: ============================================ The lag is REAL and confirmed. OTC = Binance(lag ago) × scale at 100% match. But predicting the NEXT tick's direction from this formula gives ~50% accuracy because the "noise" term is well-randomized. The developer built a system that: 1. LOOKS like a real market (tracks Binance with lag) 2. IS deterministic (coded algorithm, not random) 3. But PREVENTS exploitation (50/50 balancer + noise injection) I believe the 50/50 balancer MUST have a weakness because: - Pure randomness cannot maintain realistic market-like volatility - Pure randomness cannot guarantee 50/50 over time - The balancer itself is CODE with decision points - Those decision points are exploitable IF we can identify them ============================================ WHAT I NEED: ============================================ I need genuinely NEW approaches to break the 50/50 balancer. Not variations of what I've tried. Not "accept 55-60% and trade small." Specific questions: 1. What mathematical weakness exists in a system that must simultaneously look market-like AND maintain 50/50 balance? 2. How can the shared PRNG (same sequences across 16 assets) be exploited? 3. If the server generates all 62 assets in one 1ms batch, what constraints does that impose on individual asset directions? 4. How do real cryptographers attack constrained PRNGs? 5. What side-channel attacks exist for algorithmic price generators? 6. Is there a timing attack using the 7ms server clock cycle? 7. Can the batch generation order (which asset comes first in the batch) leak information about the PRNG state? I want approaches that are mathematically novel, not pattern-matching. The system is designed to defeat pattern-matching. I need to attack the ALGORITHM STRUCTURE, not the output patterns.