Learning When Not to Trade: A Generalized Profit-As-Regression Decision Model for Biased, Partially Observed Business Markets
P34 is a decision model, not a price forecaster. It learns profitable portfolio selection from biased historical business data where only prior accepted trades have reliable outcomes.
P34 · Difficult market
+$14.8k
Realized profit under declining margin and hidden bias
Baseline · Same market
−$219.8k
Tuned GBDT-R predicted +$126.6k but lost it all
P34 · Stationary market
+$228.6k
Matches optimistic baseline in favorable conditions
Calibration
71.7%
Prediction calibration in the difficult benchmark
PARML optimizes against final economic outcomes — not intermediate forecasts. The target is realized portfolio profit, not predicted price. The model internalizes execution risk, adverse selection, inventory loss, and margin compression by learning from what actually happened.
Businesses never observe the complete market. They only see outcomes of trades they accepted. Historical data is shaped by a prior policy the model cannot see. Standard regressors trained on these labels over-enter trades when conditions shift.
P34 constructs multiple plausible interpretations of the market, filters dominated universes, corrects for false-positive risk, and selects portfolios using meta-level telemetry. No-trade is always an available action.
Decision pipeline
The interface
The business menu: a structured set of candidate actions a business can actually take.
Request Technical ReportA business whose decisions are encoded in spreadsheets, manual judgment, and brittle rules is difficult to scale, audit, and transfer. A business whose decision process is represented by a retrainable profit-directed model becomes more stable, more observable, and more adaptive.
P34 is not a forecaster attached to a rule engine. It is an operational asset that converts business complexity into repeatable decision quality.
What current systems cannot do