Explainability and Self-Tuning

Rich structured explanations for MCP tools, adaptive confidence thresholds, and built-in backtesting.

Three subsystems that make the fund self-aware: explainability surfaces everything the LLM needs in a single call, adaptive thresholds tune decision confidence from outcomes, and the backtest runner validates strategies before deployment.

Explainability

The Explainer reads from every fund subsystem and returns structured dictionaries that give the LLM a complete picture. Each section is wrapped in error handling so a failure in one subsystem does not break the entire response.

Fund State Explanation

fund_explain()

Returns everything about the fund in one call:

Section	Contents
overview	NAV, drawdown, running strategies, kill switch status, universe size
regime	Current market regime with confidence and recent transitions
risk	Tail risk (VaR/CVaR), portfolio Greeks, correlation flags
strategies	Per-strategy status with P&L, drawdown, uptime, orders, positions
decisions	Recent autonomous decisions with reasoning and outcomes
alerts	Recent alerts with counts by category
hypotheses	Active trading hypotheses with lifecycle state and confidence
alpha_factors	Per-factor information coefficients and weights

Strategy Explanation

fund_explain_strategy(name="political_mm")

Deep dive into a single strategy:

Section	Contents
performance	NAV, P&L, drawdown, open orders, active positions, uptime
execution	Fill rates, slippage, adverse selection, market impact
alpha_decay	Current edge, half-life, predicted zero crossing, retirement recommendation
promotion	Current stage (paper/shadow/live), days in stage
hypotheses	Active hypotheses for this strategy’s markets

Risk Explanation

fund_explain_risk()

Full risk breakdown:

Section	Contents
tail_risk	Historical and Cornish-Fisher VaR/CVaR, skewness, kurtosis
greeks	Portfolio-level delta, gamma, theta, vega
correlations	Cross-strategy correlation matrix with high-correlation flags
stress	Stress test results (10/20% drawdown, correlation spike, liquidity crisis, black swan)
var_budget	Per-strategy VaR utilization vs. budget
dynamic_limits	Regime-adjusted position limits
attribution	Alpha/beta decomposition, information ratio

Full Snapshot

fund_full_snapshot()

Everything combined: status, strategies, NAV history, risk dashboard, stress test, P&L attribution, correlation matrix, execution report, hypotheses, regime, alpha model, decisions, decay report, alerts, ledger, and promotion status. Use this when the LLM needs a complete picture to make a decision.

Design

The Explainer is stateless. It holds a reference to the FundManager and reads from subsystems on demand. If a subsystem is not enabled or throws an error, that section returns "unavailable" instead of failing the entire call.

Adaptive Thresholds

Self-tuning confidence thresholds that learn from decision outcomes. Over time, the fund raises thresholds for actions that produce bad outcomes and lowers them for actions that consistently succeed.

How It Works

1. After each autonomous decision, the outcome (profitable or not) is recorded
2. Every 120 oversight ticks, the tuner computes precision (profitable / total) per action type
3. Thresholds are adjusted:

• Precision < 0.6: raise threshold by 0.05 (be more conservative)
• Precision > 0.8: lower threshold by 0.02 (allow more actions)

4. Updated thresholds are injected into the DecisionFramework

Bounds

Thresholds are clamped to [0.2, 0.95] to prevent the system from becoming either reckless or paralyzed. A minimum of 10 samples per action type is required before any adjustment.

Configuration

fund = FundManager(FundConfig(
    total_capital=100_000,
    adaptive_thresholds_enabled=True,
))

Precision Tracking

stats = fund.threshold_tuner.stats()
# {
#     "deploy": {"total": 25, "profitable": 18, "precision": 0.72},
#     "scale_up": {"total": 12, "profitable": 10, "precision": 0.83},
#     "retire": {"total": 8, "profitable": 6, "precision": 0.75},
# }

Built-in Backtester

The BacktestRunner generates synthetic prediction market data and runs strategies through it. It produces callables that plug directly into the Autopilot’s evaluate() method, so every strategy is backtested before deployment.

Synthetic Price Data

Generates bounded random walks that simulate prediction market price movement:

• Prices stay within [0.02, 0.98] (valid probability range)
• Drift is proportional to the edge estimate (positive edge = upward drift)
• Bid/ask spread and volume are included in each tick
• Default: 500 ticks per backtest run

Integration with Autopilot

When the fund is configured, the BacktestRunner provides two callables:

# These are wired automatically in the oversight loop
backtest_fn = fund.backtest_runner.as_backtest_fn()
robustness_fn = fund.backtest_runner.as_robustness_fn()

# Autopilot uses them to validate candidates before deployment
autopilot.evaluate(
    candidate,
    backtest_fn=backtest_fn,
    robustness_fn=robustness_fn,
)

The backtest function returns a results dict with sharpe, trades, max_drawdown_pct, and walk_forward_consistency. The robustness function returns a dict with p_value indicating statistical significance.

Direct Usage

from horizon.fund import BacktestRunner

runner = BacktestRunner()
results = runner.run(
    pipeline=[signal_fn, quote_fn],
    markets=["will-x-win"],
    edge=0.05,
    n_ticks=1000,
)

MCP Tools Summary

Tool	Description
fund_explain	Full fund state: overview, regime, risk, strategies, decisions, alerts, hypotheses, alpha factors
fund_explain_strategy	Single strategy deep dive: performance, execution, decay, promotion, hypotheses
fund_explain_risk	Full risk analysis: tail risk, Greeks, correlations, stress, VaR, limits, attribution
fund_hypotheses	Active trading hypotheses with lifecycle state, confidence, edge estimates
fund_regime	Current market regime with confidence and recent transitions
fund_alpha_model	Alpha model factor report: per-factor ICs, significance, weights
fund_decisions	Recent autonomous decisions with reasoning, confidence, outcome, guardrail status
fund_decay_report	Alpha decay tracking: edge erosion, half-life estimates, retirement recommendations
fund_full_snapshot	Everything in one call