Situation Room
Live intelligence feed — tiered TFI refresh every 30s for urgent layers
TFI Event Feed 0 events
Latest Updates
TFI Layer Scores
Instrument Recommendations
Active Bots —
Open Positions —
Dashboard
Open Positions
P&L Curve
Current Bias
Recent Trades
Trade Journal
Reports
P&L by Period
Instrument Distribution
Exit Reasons
R-Distribution
Summary Table
Learning & Intelligence
Layer Accuracy
Layer Weight Adjustments
Entry Pattern Performance
Instrument Ranking
Bot Status
TFI Layer States
Dynamic Weights
Bot Arena
Equity Curves (Normalized %)
Leaderboard
Forecaster Outlook
No Brainer Insights
Trade Ideas
Pattern Performance
Condition Edge
🧠 TFI
Learning Timeline
Bots
TFI Lab
Live measurement of TFI's predictive value. Per-layer hit rates, per-bot edge attribution (aligned vs counter-trend PnL), and per-instrument forecast accuracy. Empty cells until ~30 minutes after first deploy (reconciler runs every 15 min).
Per-layer hit rate (last 30 days)
Per-bot edge (aligned vs counter, last 30 days)
Per-instrument forecast accuracy (last 30 days)
Knowledge Base
Everything you need to understand how TradeForge thinks, decides, and learns. Built for humans, not quants.
What is TFI?
TradeForge Intelligence (TFI) is the brain behind every trading decision. It continuously monitors 15 independent data streams — from economic reports to central bank speeches to cross-market correlations to session timing — and combines them into a single, actionable signal for each instrument. Think of it as a team of 15 specialists, each analyzing a different dimension of the market, voting together on direction, confidence, and timing. Originally 9 layers, TFI was expanded to 15 during the Profi Traders fleet build (April 2026) to support 44 bots across diverse strategy types — from pure price action (Al Brooks) to intermarket macro (Brent Donnelly) to central bank divergence (Jarratt Davis).
How It Works — The Big Picture
Data flows through a pipeline of stages. Each stage refines the signal until it becomes a concrete trading recommendation.
The Core 9 Layers
The original 9 layers, active since March 2026. Each is an independent analyst that looks at one specific aspect of the market, outputting a score from -5 (very bearish) to +5 (very bullish). See New TFI Enhancements below for the 6 additional layers added in April 2026.
engine/surprise_index.pyPhase 2 Layers (10–15) ALL LIVE
Six new layers added in April 2026. All six are fully operational and scoring on the tiered refresh schedule (see "Tiered Refresh Architecture" below) alongside the original 9. Base weights were redistributed across all 15 layers — the system still sums to 1.0.
Tiered Refresh Architecture PHASE 18 — LIVE
Before April 7, 2026 a single 30-minute loop refreshed all 15 layers. Market-moving events (Trump tweets, BOJ interventions, RSS news) could take up to 30 minutes to influence the signal. Phase 18 split the work into two parallel loops so urgent layers refresh in 30 seconds while slower layers stay on the 30-minute cycle.
intel_signal.json through a single-writer asyncio.Lock so the bots see one consistent state.asyncio.gather so the worst-case loop time is bounded by the slowest single fetch (~800ms), not the sum.Recombine-on-Change Gate
Each loop fetches its layers in parallel, compares the new scores against the cache, and only triggers a full 15-layer recombine + signal-file write if at least one score actually changed. The fast loop polls 2,880 times per day but recombines only ~50–100 times — when scores really moved.
Stability Gate
A 2-cycle debounce on the bias output kills phantom oscillations. A flip from LEAN_LONG to LEAN_SHORT must be confirmed across two consecutive recombines before it propagates to the signal file. This prevents single-fetch noise from churning bot decisions.
Stale-While-Error Protection
Every layer holds its last-good value if the next fetch fails. Time decay then naturally reduces the influence of stale data — a 2-hour-old political score is already at ~70% strength via the 4-hour half-life. The system self-heals: a transient RSS outage doesn't crash the signal, it gradually fades out.
Per-Layer Exponential Backoff
If a layer's fetcher fails repeatedly (e.g., an RSS feed goes down), it's skipped for an exponentially growing number of cycles — capped at 60 cycles (30 minutes for the fast loop). One layer's failure never blocks the others.
Why session_killzone & central_bank are non-directional
In the Phase 1 TFI calibration (April 2026), these two layers were removed from the directional total_score because they don't represent a bullish or bearish view of the market — session_killzone is a time-of-day liquidity flag and central_bank is an intervention-risk flag. They now modify confidence and risk sizing instead. The directional layer count dropped from 14 to 12, so confidence thresholds were rescaled from 9/7/5/3 to 8/6/4/3, and the alignment bonus from ≥6 to ≥5.
Equity Prior +0.4 for US Indices
SP500, Nasdaq, and US30 receive a +0.4 score prior on every recombine to break the structural "never bullish" ceiling caused by political/sentiment layers being persistently negative during high-noise periods. The prior is small enough that a real bear signal still wins, but large enough to flip LEAN_SHORT → NEUTRAL → LEAN_LONG in calm regimes.
How Layers Combine
Each layer's score is multiplied by its weight. The weights determine how much influence each layer has on the final decision. Heavier weight = more influence.
Although Political has a 0.10 base weight, it is excluded from the directional calculation. Its weight only affects confidence reduction. The other 14 layers decide direction; Political only decides how sure we are about it.
Dynamic Weights — The Learning Loop
Weights aren't fixed. TFI adjusts them in real-time based on two mechanisms:
The final learned weight per layer = accuracy_mult × quality_mult. Accuracy measures whether the layer predicted the correct direction. Quality (R-capture) measures how much reward was actually captured when the layer was active — a layer can be directionally right but consistently enter late or exit early. Both multipliers require 30+ trades before activating. Quality mult uses a 60-day half-life so stale performance gradually fades.
Extreme Amplification
When a layer detects something extreme (score of 4 or 5), its weight temporarily increases to make that signal louder:
| Score ±5 | 2.5x weight | Maximum alarm |
| Score ±4 | 2.0x weight | Strong alarm |
| Score ±3 | 1.5x weight | Moderate alarm |
| Score ±1-2 | 1.0x weight | Normal |
Accuracy Mult (after 30+ trades)
TFI tracks which layers predicted the correct direction. Accurate layers get permanently boosted:
| Accuracy >65% | 1.5x multiplier | Trusted layer |
| Accuracy >55% | 1.2x multiplier | Good layer |
| Accuracy 45-55% | 1.0x multiplier | Neutral |
| Accuracy <45% | 0.6x multiplier | Demoted |
| Accuracy <35% | 0.3x multiplier | Nearly muted |
Quality Mult — R-Capture (after 30+ trades)
TFI tracks how much available reward was captured on trades where a layer was active and agreed with direction. Attribution weighted by signal strength (±5 = full credit, ±1 = 20%) and 60-day half-life decay:
| Avg quality >0.75 | 1.3–1.5× mult | Excellent execution |
| Avg quality 0.5–0.75 | 1.1–1.3× mult | Good execution |
| Avg quality ~0.5 | 1.1× mult | Neutral |
| Avg quality 0.25–0.5 | 0.9–1.0× mult | Poor execution |
| Avg quality <0.25 | 0.7–0.9× mult | Very poor execution |
After 200+ trades, TFI discovered that the Econ Surprise layer was the most accurate predictor. The learning loop automatically boosted its weight from 0.12 to 0.17 — a 42% increase — while demoting less accurate layers like Macro and Calendar. If Econ Surprise also produces high R-capture trades (quality mult 1.3×), its combined multiplier becomes 1.5 × 1.3 = 1.95×, nearly doubling its influence.
Time Decay — Freshness Matters
Not all data stays relevant. Technical signals from 1 hour ago matter more than from 24 hours ago. Each layer has its own "half-life" — the time it takes for a signal to lose half its strength.
Vetoes — Safety First
Before any signal reaches a bot, it passes through safety checks that can override everything. If conditions are too dangerous, TFI forces a HOLD regardless of what the layers say.
Market in crisis mode. All trading paused — forced NEUTRAL.
Fear & Greed >85, Put/Call <0.5, VIX <12 — too calm, reversal imminent.
NFP, FOMC, or CPI day — HOLD equities until the data drops.
News & Geopolitical Monitoring
TFI scans RSS news feeds every 15 minutes, looking for market-moving headlines. It uses keyword scoring to evaluate impact.
Bearish Keywords (push score down)
Bullish Keywords (push score up)
Central Bank Intervention Zones
TFI monitors when currency prices approach levels where central banks have historically intervened. When risk is HIGH, FX recommendations are overridden.
| Central Bank | Pair | Danger Zone | What They Do |
|---|---|---|---|
| 🇯🇵 BOJ | USDJPY | 160+ (extreme) | Sell USD to defend yen (acted at 161.9 in 2024) |
| 🇨🇭 SNB | USDCHF | 0.84 or below | Buy EUR/CHF to defend franc floor |
| 🇨🇳 PBOC | USDCNY | 7.35+ (extreme) | Sell USD to defend yuan |
| 🇬🇧 BOE | GBPUSD | 1.05 or below | Buy GBP to defend pound |
| 🇪🇺 ECB | EURUSD | 0.95 (parity) | Buy EUR to defend euro |
Per-Instrument Scoring
Different instruments care about different things. Gold responds to fear, oil responds to geopolitics, Bitcoin responds to risk appetite. TFI uses custom weight profiles for each.
| Instrument | Primary Driver | Secondary | Weight Mix |
|---|---|---|---|
| S&P 500 / NASDAQ / US30 | Global TFI Score | Technical + Geo | 70% global 15% tech 15% geo |
| Gold | Inverted Sentiment (fear = bullish) | COT + Geopolitical | 30% sent 25% COT 20% geo |
| Oil | Geopolitical Risk | COT + Technical | 25% geo 25% COT 20% tech |
| Bitcoin | Risk-On Sentiment | Technical + Cross-Asset | 30% sent 25% tech 20% cross |
| FX Pairs (EUR, GBP, JPY...) | Dedicated FX Outlook Engine | Geo + Econ Surprise | 80% FX engine 10% geo 10% econ |
For equities, high fear (VIX) is bearish. But for gold, high fear is bullish — people buy gold as a safe haven when they're scared. TFI automatically flips the sentiment score for gold.
From Score to Trade
The final score is converted to a human-readable bias and a confidence level, which determines how aggressively bots can trade.
Score Scale
< -2.0
-2.0 to -0.5
-0.5 to +0.5
+0.5 to +2.0
> +2.0
Confidence Levels
What boosts confidence?
Calendar & Structural Windows
Certain times of the month, quarter, or year have predictable market behavior patterns that TFI accounts for.
Quarter-End (Mar, Jun, Sep, Dec)
Last 5 trading days of each quarter. Fund managers "window dress" their portfolios — buying winners and selling losers to look good in reports. This creates unusual price movements that aren't driven by fundamentals.
Month-End Rebalancing
Pension funds and index trackers rebalance at month-end. If stocks outperformed bonds, they sell stocks and buy bonds to return to target allocations. This creates predictable flows in the last 2-3 days.
Turn-of-Month Effect
Last trading day + first 3 days of the new month tend to be bullish for stocks. This is when salary contributions flow into retirement accounts and 401(k)s, creating steady buying pressure.
High-Impact Event Days
On FOMC (8x/year), NFP (monthly), and CPI (monthly) days, TFI forces a HOLD on equities until the data is released. Trading before major data is gambling, not trading.
Signal Alignment Bonus
When multiple layers independently agree on the same direction, TFI amplifies the signal. Consensus = conviction.
If Sentiment says bearish, Flows says bearish, and Cross-Asset confirms risk-off — that's 3 completely independent data sources agreeing. This is much more significant than one layer screaming alone. The 1.3x bonus rewards this convergence.
Two Time Horizons
TFI produces two separate signals: one for short-term trading (hours) and one for longer-term positioning (weeks).
Short-Term (1-24 hours)
Uses fast-moving layers: Technical, Political, Econ Surprise, Calendar. These change rapidly and are best for intraday or next-day decisions.
Long-Term (1-4 weeks)
Uses slow-moving layers: Macro, Flows, Sentiment, Institutional, Cross-Asset. These represent deeper market currents that take time to play out.
Report History
Development changes, Trading Advice Board findings, and system improvements.
COT Reports
CFTC Commitments of Traders — institutional positioning intelligence
BOJ Intervention Monitor
5-stage early warning system for Bank of Japan currency intervention
This BOJ Monitor Section was done based on: click to expand
Advice Trading Board
A group of specialists who review, discuss, plan, and drive the TradeForge system forward together.
Project Roadmap
TradeForge development phases — from virtual trading to live