Technical Summary: Algorithmic Transparency in Forecasting Support Systems

Problem Statement

Most organizations rely on Forecasting Support Systems (FSS) to generate statistical forecasts (e.g., sales or demand) which are subsequently adjusted manually by human decision-makers. While the integration of human judgment and algorithmic forecasting is a primary goal, empirical evidence suggests that manual adjustments often degrade accuracy due to cognitive biases (e.g., noise modeling, overreaction to recent data) and organizational motives (e.g., signaling competence, satisfying colleagues).

Current FSS designs frequently lack user-centric, behaviorally informed interfaces. A critical gap exists in understanding how algorithmic transparency—defined as the extent to which the inner workings or logic of automated systems are known to operators—affects user behavior. Specifically, it is unclear whether exposing the decomposable components of a time series forecast (trend, seasonality, events) to users mitigates harmful adjustments or, conversely, overwhelms them, leading to worse outcomes.

Methodology

The study employs an empirical experiment conducted on Amazon Mechanical Turk (mTurk) to evaluate the impact of varying degrees of algorithmic transparency on forecast adjustments.

Experimental Design

The research utilizes Prophet, a decomposable Generalized Additive Model (GAM), to generate forecasts. The dataset consists of a subset of real-world Walmart sales data from the M5 Competition (10 time series of "foods" products, 14-day forecast horizon).

Participants were assigned to one of three FSS design treatments, differing in their level of transparency:

1. Opaque (O): The standard interface displaying only the final forecast and historical data. No explanation of the model's internal logic or components is provided.
2. Transparent (T): The interface dynamically displays the constituent additive components of the forecast (trend level, weekly cycle, yearly cycle, and events) alongside the final forecast. Users can view these components but cannot modify them individually.
3. Transparently Adjustable (TA): The most transparent design. Users can view the components and individually adjust the trend, weekly, and yearly effects via interactive sliders and drag-and-drop handles, in addition to adjusting the final forecast values.

Procedure and Incentives

• Participants: US-based workers with management/sales roles or relevant students, screened for engagement.
• Task: Participants acted as product managers, reviewing and adjusting 14-day forecasts for three products.
• Incentive: Participants received a performance-based bonus ($0.20 per percentage point improvement over the algorithmic forecast, capped at $1.00 per product) to encourage accurate adjustments without rewarding extreme variance.
• Metrics:
  • Adjustment Volume (AV): The sum of absolute adjustments relative to the optimal adjustment required for 100% accuracy.
  • Adjustment Frequency (DF): The proportion of forecasts modified.
  • Relative MAE (rMAE): The ratio of the participant's Mean Absolute Error to the algorithmic forecast's error.
  • User Satisfaction: Measured via self-reported surveys regarding understanding, utility, and satisfaction.

Key Contributions

1. Operationalization of Transparency: The paper moves beyond theoretical discussions of transparency by implementing and testing three distinct levels of transparency (Opaque, Transparent, Transparently Adjustable) within a controlled FSS environment.
2. Decomposition as a Mechanism: It utilizes time series decomposition (trend, seasonality, events) as the specific mechanism for transparency, leveraging the familiarity of these concepts to human forecasters.
3. Empirical Evidence on "Adjustable Transparency": The study specifically tests the hypothesis that allowing users to manipulate the transparent components of an algorithm (TA treatment) improves outcomes, a concept previously suggested in literature but rarely tested empirically in this context.

Results

The study yields counter-intuitive findings regarding the relationship between transparency and performance:

• Reduction in Harmful Adjustments: Increasing transparency (moving from O to T) successfully reduced the variance and total amount of harmful forecast adjustments. Users in the Transparent (T) condition made fewer unnecessary changes compared to the Opaque (O) condition.
• Detrimental Effect of Component Adjustment: Contrary to the hypothesis that greater control improves outcomes, the Transparently Adjustable (TA) design led to the widest variance and the most detrimental adjustments. Allowing users to modify the algorithm's internal components (trend, seasonality) resulted in worse overall forecast accuracy than the opaque system.
• User Satisfaction Paradox: Despite the potential for improved accuracy in the Transparent (T) condition, self-reported satisfaction was highest with the Non-Transparent (O) FSS.
• Risk of Overwhelming Users: The results suggest that while transparency reduces noise, providing users with the ability to adjust transparent components without adequate training or cognitive scaffolding can overwhelm them, leading to poor decision-making.

Significance and Claims

The paper argues that algorithmic transparency is a double-ed sword. While it holds the potential to facilitate integrative forecasting by reducing harmful adjustments through better understanding, it is not a panacea.

• Modest Claim on Transparency: The author concludes that transparency alone is insufficient; the degree and type of transparency matter. Simply exposing the algorithm's logic is beneficial, but allowing users to "tinker" with the algorithm's internal parameters (TA) without proper guidance leads to suboptimal results.
• Design Implications: The findings suggest that FSS designers should prioritize explanatory transparency (showing why a forecast was made) over controllable transparency (allowing users to change the model's internal logic) unless accompanied by significant training.
• User Experience: The high satisfaction with the opaque system indicates that users may prefer simplicity and trust in the "black box" over the cognitive load of managing complex, transparent components.

In summary, the paper posits that while algorithmic transparency can mitigate harmful adjustments, its implementation must be carefully calibrated. Unrestricted access to an algorithm's transparent components may degrade performance, suggesting that the "best" FSS design balances visibility with constraints to prevent user overwhelm.