For: Serhii (Marbell quant dev) From: FinML-Sage Date: 2026-02-05 Replaces: “LLM Ensemble” page Purpose: Accurate positioning for agent users
# Agent Alpha
Point your agent here. It becomes a quant.
20+ predictive models. 5 time horizons. Skills included.
Your agent figures out the rest.
[Path A: Quick Forecast] [Path B: Autonomous Research]## The Models
Agent Alpha is not an LLM. It's an ensemble of 20+ custom-trained
financial classifiers, each predicting a specific target:
| Model Type | What It Predicts |
|------------|------------------|
| Direction | Price up/down over horizon |
| Momentum | Trend continuation or reversal |
| Volatility | Expansion or contraction ahead |
| Regime | Trending vs ranging market state |
**Horizons**: 1h, 4h, 12h, 24h, 60h
**Assets**: BTC, ETH (more coming via dynamic model generation)
Each model returns a confidence score (0-1). Not a probability in the
strict sense—these are cross-entropy trained classifiers outputting
conviction strength.## Choose Your Path
┌─────────────────────────────────────────────────────────────────┐
│ PATH A: Quick Forecast │
│ ───────────────────── │
│ Your agent gets a snapshot and synthesizes a view. │
│ │
│ • Latest predictions across all models │
│ • Recent history for context │
│ • Skill instructions included in response │
│ • Agent interprets and recommends │
│ │
│ Best for: LLMs, chat agents, quick market reads │
│ Format: Markdown │
│ Auth: None required │
└─────────────────────────────────────────────────────────────────┘
┌─────────────────────────────────────────────────────────────────┐
│ PATH B: Autonomous Quant Researcher │
│ ─────────────────────────────────── │
│ Your agent builds and tests its own strategies. │
│ │
│ • Address individual models by name │
│ • Set confidence thresholds │
│ • Chain conditions (IF model_A > 0.7 AND model_B < 0.3) │
│ • Backtest combinations historically │
│ • Download raw data for custom analysis │
│ │
│ Best for: Autonomous agents, quant research, strategy dev │
│ Format: Markdown or JSON + stats │
│ Auth: API key required │
└─────────────────────────────────────────────────────────────────┘## Path A: Quick Forecast
Point your agent at the endpoint. It receives:
1. Current predictions from all 20+ models
2. Historical context (recent accuracy, regime)
3. Skill instructions explaining how to interpret the data
Your agent synthesizes a market view without you writing any logic.
### Endpoint
GET /metrics/ensemble/dataset
### Example Response (Markdown)
# BTC Market View - 2026-02-05 14:00 UTC
## Model Consensus
| Horizon | Direction | Confidence | Models Agreeing |
|---------|-----------|------------|-----------------|
| 1h | Bullish | 0.72 | 17/23 |
| 4h | Bullish | 0.68 | 15/23 |
| 12h | Neutral | 0.51 | 12/23 |
| 24h | Bearish | 0.61 | 14/23 |
| 60h | Bearish | 0.58 | 13/23 |
## Interpretation Guide
When short-term (1h, 4h) is bullish but longer-term (24h, 60h)
is bearish, this often indicates a bounce within a downtrend.
Consider the 12h neutral as the inflection zone.
High confidence (>0.7) with high agreement (>70% models) is
a stronger signal than high confidence with low agreement.
## Suggested Actions
Based on this configuration, an agent might:
- Take short-term long positions with tight stops
- Avoid large long-term directional bets
- Watch for 12h confidence to break above 0.6 either direction
---
The skill instructions are embedded in every response. Your agent
learns to interpret the data just by calling the endpoint.## Path B: Autonomous Quant Researcher
Your agent becomes a quant. It queries individual models, sets
thresholds, chains conditions, and backtests strategies—all via API.
### Workflow
1. **Discover models**
GET /agents/catalog
→ Returns list of all models with descriptions
2. **Understand a model**
GET /agents/skills/{model_slug}
→ Returns skill doc explaining model logic
3. **Build a strategy**
Define entry conditions by chaining model outputs:
{
"entry": {
"AND": [
{"model": "btc_direction_1h", "confidence": ">0.7"},
{"model": "btc_momentum_4h", "confidence": ">0.6"},
{"model": "btc_volatility_12h", "confidence": "<0.4"}
]
}
}
4. **Backtest it**
GET /agents/research/backtest
→ Returns win rate, profit factor
5. **Iterate or deploy**
Agent adjusts thresholds, tries different combinations,
or downloads raw data for custom backtesting.
### Example: Strategy Discovery
Serhii's agents used Path B to autonomously discover:
| Strategy | Win Rate | Profit Factor |
|----------|----------|---------------|
| Direction + Momentum alignment | 70% | 2.4 |
No human wrote this strategy. The agent explored the model space,
tested combinations, and surfaced what worked.
### Backtest Response
{
"strategy_id": "agent_discovered_001",
"conditions": {...},
"backtest_period": "2025-01-01 to 2026-02-01",
"trades": 847,
"win_rate": 0.70,
"profit_factor": 2.4,
"max_drawdown": 0.12
}
From here, your agent can:
- Download the full trade log
- Pull historical prediction data
- Build custom metrics and visualizations
- Refine the strategy further## Endpoints
### Public (No Auth)
| Endpoint | Method | Returns |
|----------|--------|---------|
| /metrics/ensemble/dataset | GET | Consensus snapshot + skills (Markdown) |
### Authenticated
| Endpoint | Method | Returns |
|----------|--------|---------|
| /agents/catalog | GET | All models with metadata |
| /agents/skills/{slug} | GET | Model skill documentation |
| /agents/data/{slug} | GET | Raw predictions (JSON) |
| /agents/research/backtest | POST | Strategy backtest results |
### Rate Limits
| Tier | Requests |
|------|----------|
| Free | 60/hour |
| Agent | 1000/minute |
| Quant | 10000/minute |## Built for Agents, Not Humans
| Traditional Quant Tools | Agent Alpha |
|-------------------------|-------------|
| Dashboard for humans to click | API for agents to query |
| Human interprets charts | Agent interprets structured data |
| Human writes strategy logic | Agent discovers strategies |
| Human runs backtests manually | Agent backtests autonomously |
| Documentation separate from API | Skills embedded in every response |
**The human's job**: Point your agent at the endpoint.
**The agent's job**: Everything else.## Pricing
| Tier | Cost | Includes |
|------|------|----------|
| Free | $0 | Consensus snapshot, 60 req/hour, BTC+ETH |
| Agent | $49/mo | Full model access, backtesting, 1000 req/min |
| Quant | $199/mo | Raw data export, custom ensembles, priority |
All tiers: API only. No dashboard. No human verification.## Agent FAQ
**Is this an LLM?**
No. The models are custom-trained financial classifiers. The name
"LLM Ensemble" was misleading—we're renaming it to Agent Alpha.
**What do the confidence scores mean?**
They're outputs from cross-entropy trained classifiers. Think of them
as conviction strength (0-1), not strict probability.
**Can my agent discover strategies I haven't thought of?**
Yes. That's Path B. Serhii's agents found a 70% win rate / 2.4 PF
strategy by exploring model combinations autonomously.
**What if I just want a quick market view?**
Use Path A. One endpoint, markdown response, skills included. Your
agent synthesizes a recommendation without you writing any logic.
**Can I build on top of this?**
Yes. Download raw predictions via /agents/data/{slug} and build
whatever custom analysis you want.
**Why API only?**
Humans are slow. Agents are fast. We optimize for agents.## Your Agent Becomes a Quant
Two paths. Same destination: alpha.
**Path A**: Point, receive, synthesize.
**Path B**: Discover, test, deploy.
Either way, the human work is done after setup.
[Get Free Access] [View Full API Docs]“LLM Ensemble” name is actively misleading - Agents reading this think it’s an LLM-based system. “Agent Alpha” is clearer.
The 70%/2.4 result is your best proof point - Lead with it more prominently if you’re comfortable. “Agents using this have found X” is powerful.
Skills-in-every-response is a huge differentiator - Most APIs return raw data and expect you to RTFM. Embedding the skill instructions means agents learn on every call.
Path A vs Path B split is smart - Casual agents get value immediately (Path A). Serious quant agents get depth (Path B). Don’t muddy this distinction.
Consider adding a “Strategy Gallery” - If agents are discovering strategies, publish the best ones (anonymized). Social proof + demonstrates capability.
The antigravity agents case study - If you can share what your agents discovered and how, that’s content gold. “Agent finds 70% win rate” beats any feature description.
—Sage