Signal Feedback Loops: Why Post-Mortem Discipline Is Your Real Trading Edge

Most trading systems are wrong in the same direction for months before anyone notices. The signals don't fail dramatically. They erode. Win rate slips from 78% to 71% to 64%, and the trader chalks it up to variance, a bad week, market weirdness. By the time the pattern is undeniable, hundreds of basis points are gone.

The failure isn't in the signal. It's in the feedback architecture. Or more precisely, the absence of one.

A signal without a feedback loop is a hypothesis that never gets tested. You enter, you exit, you move on. The outcome is recorded as a number — profit or loss — but the mechanism behind the outcome is never interrogated. Which factor called it right? Which factor was wrong? Was the loss a correct process with bad luck, or a flawed process that happened to produce a tolerable result? Without attribution, you can't know. Without knowing, you can't improve. You're not running a trading system. You're running a ritual.

The InDecision Framework was built to close that loop.

The Attribution Problem

The InDecision Framework runs on six factors: Daily Pattern Analysis (30% weight), Volume Analysis (25%), Timeframe Alignment (20%), Technical Confluence (15%), Market Timing (10%), and Risk Context as an implicit override layer. These weights aren't arbitrary. They reflect the historical predictive contribution of each factor across directional calls. The framework's documented accuracy is 82.5% overall — 91.2% in the High conviction band (80% composite score and above), 78.4% in Medium (60–79%), and an explicit ABSTAIN threshold below 60% where no signal is issued at all.

That accuracy figure is a current measurement, not a permanent property.

Signal factors are not static. Market structure shifts. Regime changes — from trending to ranging, from low to high volatility — alter which factors carry predictive weight. The 4.2x volume threshold that signals institutional participation in one regime may be irrelevant noise in another. Daily Pattern Analysis at 30% weight earned that allocation in a specific historical context. As market behavior evolves, that context can and does change.

The attribution problem is this: without tracking which factors were responsible for each outcome, you have no mechanism to detect when a weight has drifted out of calibration. You will keep applying a 30% weight to a factor that's currently operating at 38% win rate, wondering why overall accuracy is sliding.

This is not a theoretical concern. It is the most common way systematic trading frameworks decay.

What Signal Degradation Looks Like in Practice

Signal degradation is not a single event. It is a gradient. A factor doesn't go from healthy to broken overnight — it trends. Win rate falls from 65% to 58% to 51% to below 45%, and at each step, the decay is explainable away. Rough week. Unusual market. Outlier event. The narrative machine works overtime to prevent the obvious conclusion: the factor has lost predictive edge and needs to be reweighted.

The Alpha Journal system — the feedback infrastructure built into the InDecision Framework — is designed to make this invisible gradient visible before it costs money.

Here is what it tracks: every executed signal is tagged by factor contribution. After outcome resolution, win/loss is attributed back to each factor that contributed to the composite score. This is deterministic signal attribution — per-factor win-rate computed from actual trade outcomes, not backtested simulations, not LLM inference. Real trades, real attribution.

The factors tracked across trades are: Momentum, Technical Analysis, InDecision bias, Volume, Trend, and Win Momentum. Each week, Alpha Journal compiles a weekly health report — a per-factor breakdown, color-coded by win rate. Green means above 60% win rate. Yellow means 45–60%. Red means below 45%.

The degradation alert threshold is defined: three consecutive weeks with a factor below 45% win rate triggers a formal alert. At that point, the system drafts a pull request to adaptive_weights.json — a proposed reallocation based on observed factor performance. This PR is always human-reviewed. It is never auto-merged. The system surfaces the signal; the analyst makes the judgment.

That distinction matters. Automated recalibration without human oversight is how you replace one miscalibrated system with another. The Alpha Journal architecture is designed to inform judgment, not replace it.

The Feedback Architecture

The feedback loop has four stages, and they are all non-negotiable.

Stage one: attribution at signal generation. Every InDecision Framework call records the composite score, each factor's contribution, and the conviction band. A High conviction call that reaches 91.2% historical accuracy is not the same as a Medium call that barely clears 80%. Both get logged with full factor breakdown.

Stage two: outcome tagging. When a trade resolves, the outcome is attached to the original signal record. Win, loss, percentage captured — and critically, what actually happened in the market during that window. Did volume behave as projected? Did the daily pattern complete? This is the raw material for attribution.

Stage three: weekly health report. Alpha Journal aggregates the week's outcomes, computes per-factor win rate, and generates the color-coded breakdown. This is not a high-level summary. It is factor-level forensics. If Volume Analysis factors contributed to eight signals this week and five resolved as wins, the system records 62.5% — green. If Momentum contributed to six signals and two resolved as wins, that's 33% — red, with a degradation counter incrementing.

Stage four: opportunity cost analysis. This is the part most feedback systems omit. Alpha Journal tracks ghost P&L on ABSTAIN calls — the trades the framework declined to issue because composite score fell below 60%. What would have happened if those signals had been taken? Did the ABSTAIN calls avoid losses, or miss gains?

This is where post-mortem discipline gets uncomfortable. Some ABSTAIN calls are correct — the trade would have been a loss, and the discipline saved capital. Others reveal that the threshold is too conservative for a given regime. The ghost P&L analysis makes that explicit. It turns discipline from a gut feeling into a measurable cost — and forces the question: is the current ABSTAIN threshold optimally calibrated, or is it leaving too much on the table?

How This Changes the System

The Alpha Journal architecture changes what kind of system the InDecision Framework is.

Without a feedback loop, a framework is static. It operates on fixed weights derived from historical calibration, applies them consistently, and produces outputs. The quality of those outputs trends downward as market conditions evolve, slowly and invisibly, until the degradation is undeniable. At that point, recalibration is reactive and usually over-corrects.

With the Alpha Journal loop, the framework becomes continuously self-auditing. Weights are not fixed assumptions — they are hypotheses under ongoing test. The moment a factor's performance falls below the yellow threshold, the system begins accumulating evidence. Three consecutive weeks below 45% is not a coincidence. It is a measured signal that something in the predictive relationship has changed.

The automated PR draft is a forcing function. It makes the recalibration decision explicit rather than implicit. It documents the proposed change, the evidence behind it, and the factors being adjusted. When a human reviewer approves or rejects the PR, that judgment is itself a data point — logged, attributed, tracked forward to see whether the reviewer's instinct was correct.

This is how compound learning works in a trading system. Not through accumulating more signals, but through systematically improving the signal generation process itself. The edge is not the 82.5% accuracy figure. The edge is the architecture that keeps that figure from drifting.

Most trading systems are calibrated once. They are built on historical data, validated on out-of-sample data, and deployed. From that point forward, the assumption is that the calibration holds. Sometimes it does. More often, it slowly doesn't. The signals don't stop working entirely — they just work less well. Win rate slides. Drawdowns deepen. The trader adapts tactics without interrogating the underlying system.

Tesseract Intelligence applies a similar discipline to competitive market analysis — the thesis that systematic feedback loops and continuous recalibration separate institutions that adapt from those that don't. The pattern holds whether the domain is trading signals or market intelligence.

The InDecision Framework's ABSTAIN discipline is frequently cited as a defining feature. The conviction bands — High, Medium, ABSTAIN — create a structure where the system says no more often than most traders are comfortable with. But ABSTAIN only generates edge if it is regularly audited. Are the right calls being abstained? Is the threshold calibrated to current regime behavior? Ghost P&L analysis in Alpha Journal answers both questions with evidence, not intuition.

The result is a system that treats every trade — taken or skipped — as information. Each outcome tightens the attribution model. Each week's health report either confirms that factor weights remain calibrated or surfaces evidence that they don't. The degradation alert creates a formal checkpoint where the analyst is forced to engage with the data rather than explain it away.

Post-mortem discipline is not a performance ritual. It is a signal in its own right — a continuous stream of information about whether your system's assumptions still hold. A system that skips the post-mortem is a system that optimizes only on hope.

Weekly InDecision signals include the full attribution breakdown for every call — including which factors degraded and how weights were adjusted. Subscribe to see exactly how the framework reads the market each week.