Aligning Financial Forecasting with Trading Utility

This direction is merged into another direction: metric-utility misalignment. Its last version stays readable and citable here.

Version 1 of 114 papers24 conceptsfrozen 2026-10-04

In traditional machine learning, the goal is usually to minimize the distance between a prediction and the ground truth. If a stock is priced at \$100 and a model predicts \$101, a standard loss function like Mean Squared Error (MSE) treats this as a minor success. However, if the actual price drops to \$99, that "minor" error is a catastrophic failure for a trader: the model signaled a "buy" when it should have signaled a "sell."

A growing body of research suggests that the "cognitive dissonance" identified by Fabio Cuzzolin and Wojtek Buczynski — where academic models claim high accuracy but fail in real-world funds — stems from this fundamental misalignment. To bridge this gap, researchers are moving away from point-in-time price regression toward models that prioritize directional consistency and cumulative financial utility.

The Failure of Distance-Based Metrics

The core problem is that standard statistical metrics like MSE, RMSE, and MAPE are "direction-blind." As Cristian Tiu and Yi Wei argue, conventional regression treats prediction as a task of minimizing distance, which ignores the binary nature of market movement.

Consider two models:

  1. Model A predicts a \$2.00 increase; the price actually increases by \$0.10.
  2. Model B predicts a \$0.05 decrease; the price actually increases by \$0.10.

Statistically, Model B has a lower error. Practically, Model A is the only one that makes money. Han Li notes that models optimized for precise price points frequently fail to predict whether the market will move up or down. This discrepancy is so pronounced that Hongkang Zhang and Tiantian Gao developed FinTSBridge, an evaluation suite that uses correlation-aware metrics to distinguish between models that actually capture trends and those that simply predict "no change" to minimize MSE.

Prioritizing the Sign Over the Magnitude

If the goal is trading, the sign of the change (positive or negative) is often more important than the magnitude. Yong Shi and Wen Long demonstrate this with GP-LSTM, a model that uses deep kernels to identify the sign of market movements. They found that prioritizing directional accuracy over absolute magnitude error leads to superior trading outcomes, especially during market turbulence.

Similarly, Zhiyang Chen observed a distinct trade-off: while LSTMs might be better at absolute price accuracy, Transformer-based architectures often excel at predicting the direction of movement. This suggests that the "best" model depends entirely on whether you are measuring statistical fit or strategic utility.

Direct Optimization for Profit

Rather than training a model to predict prices and then building a strategy on top of those predictions, a more direct approach is to make profit the objective function itself.

  1. Reinforcement Learning (RL): Ramit Sawhney and Arnav Wadhwa developed PROFIT, a DRL framework that bypasses price prediction entirely. Instead, it learns "Buy, Sell, or Hold" actions by maximizing a reward function based on actual portfolio returns and risk management.
  2. Profit-Guided Loss Functions: Devroop Kar and Alexander G. Ororbia propose training standard neural networks using loss functions that directly optimize for financial returns. By aligning the gradient descent process with profit rather than error, their Crossformer model outperformed traditional "buy and hold" strategies even in bear markets.

Visual and Trajectory-Based Evaluation

Because numerical error metrics can be deceptive, researchers like Minjong Cheon and Hanseon Joo advocate for visual trend analysis and trajectory tracking. They argue that a model with high RMSE might still be highly valuable if it correctly follows the general movement of the market. This shift in perspective moves the field away from "point-wise" accuracy toward "path-wise" utility, ensuring that machine learning models serve the actual needs of investors rather than just minimizing a mathematical abstraction.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_aligning_financial_forecasting_with_trad_131b6435,
  title = {Aligning Financial Forecasting with Trading Utility},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/aligning-financial-forecasting-with-trading-utility/rdr_131b6435dd434cc2925984da9ad10ce3},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_aligning_financial_forecasting_with_trad_131b6435_v1,
  title = {Aligning Financial Forecasting with Trading Utility},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/aligning-financial-forecasting-with-trading-utility/rdr_131b6435dd434cc2925984da9ad10ce3/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}