Carlos

• Updated: January 30, 2026
• 7 min read

Modeling Cascaded Delay Feedback for Online Net Conversion Rate Prediction: Benchmark, Insights and Solutions

Direct Answer

The paper introduces TESLA, a novel framework for predicting net conversion rates (NetCVR) in e‑commerce by explicitly modeling the cascaded delay feedback of clicks, conversions, and refunds. By jointly learning these interdependent events, TESLA delivers more accurate revenue‑focused forecasts, enabling businesses to allocate marketing spend and inventory with higher confidence.

Background: Why This Problem Is Hard

Online retailers and advertisers rely on conversion‑rate predictions to drive bidding strategies, inventory planning, and user‑experience personalization. Traditional CVR models focus on the immediate probability that a click leads to a purchase, ignoring two critical realities:

• Delayed feedback: Purchases often occur hours or days after the initial click, creating a temporal gap that hampers timely model updates.
• Post‑purchase churn: Refunds, cancellations, and chargebacks retroactively reduce the net revenue generated by a conversion, a factor rarely captured in standard pipelines.

Existing approaches attempt to address delay by applying survival analysis or by treating refunds as a separate binary outcome. However, these methods suffer from two major limitations:

1. Fragmented learning: Modeling clicks, conversions, and refunds independently discards the causal chain linking them, leading to biased estimates.
2. Data sparsity: Refund events are rare compared to clicks, making it difficult for isolated models to learn reliable patterns without overfitting.

Consequently, businesses either over‑estimate revenue (by ignoring refunds) or under‑react to delayed conversions (by using stale data), both of which erode profitability.

What the Researchers Propose

To overcome these challenges, the authors propose the TESLA (Temporal Event Sequence Learning Architecture) framework, which treats the click‑conversion‑refund pipeline as a single, unified stochastic process. The key ideas are:

• Joint modeling: A shared representation learns from all three event types simultaneously, allowing information from abundant click data to regularize the scarce refund signal.
• Temporal attention: The model dynamically weighs historical interactions based on their elapsed time, capturing the decay of relevance for older clicks.
• Counterfactual correction: By estimating the probability of a refund conditioned on the conversion, TESLA adjusts the raw conversion forecast to produce a net conversion rate that reflects expected revenue loss.

The framework consists of three logical agents:

1. Click Encoder – ingests raw click logs, extracts user‑item features, and produces a time‑aware embedding.
2. Conversion Predictor – consumes the click embedding and predicts the likelihood of a purchase within a configurable horizon.
3. Refund Adjuster – conditions on the predicted conversion and historical refund patterns to estimate the expected refund probability, yielding NetCVR.

How It Works in Practice

At a high level, TESLA operates as a pipeline that processes streaming event data in near‑real time. The workflow can be broken down into four stages:

1. Data Ingestion and Pre‑processing

Click, conversion, and refund logs are unified into a single event stream. Each event is timestamped and enriched with contextual features (device type, campaign ID, product category, etc.). Missing timestamps for refunds are back‑filled using the original conversion time.

2. Temporal Embedding Generation

The Click Encoder applies a position‑aware transformer that incorporates elapsed time as a bias term. This yields a sequence of embeddings where recent clicks receive higher attention weights, while older clicks gradually fade.

3. Joint Prediction

The Conversion Predictor and Refund Adjuster share the same backbone network but diverge at task‑specific heads. The conversion head outputs P(conversion | click, t), while the refund head outputs P(refund | conversion, t). Because both heads draw from the same embedding, learning is synergistic.

4. NetCVR Computation

NetCVR is derived by multiplying the conversion probability with the complement of the refund probability:

NetCVR = P(conversion | click) × (1 − P(refund | conversion))

This simple arithmetic masks the underlying complexity, delivering a single, revenue‑oriented metric that can be fed directly into bidding engines or inventory allocation models.

What distinguishes TESLA from prior work is its end‑to‑end training regime. Instead of training three separate models and stitching their outputs, TESLA optimizes a unified loss that balances conversion accuracy against refund calibration, ensuring that improvements in one sub‑task do not degrade the other.

Evaluation & Results

The authors evaluate TESLA on two large‑scale e‑commerce datasets, including the newly released CASCADE dataset that contains over 200 million click‑conversion‑refund triples across multiple product categories. Evaluation focuses on three dimensions:

• Predictive performance: Area under the ROC curve (AUC) for conversion and refund predictions, and mean absolute error (MAE) for NetCVR.
• Temporal robustness: Performance degradation when the model is applied to data with varying delay windows (e.g., 1‑day vs. 7‑day conversion windows).
• Business impact: Simulated bidding experiments measuring revenue lift when using TESLA‑derived NetCVR versus traditional CVR.

Key findings include:

In the bidding simulation, TESLA’s NetCVR led to a 5.8% increase in advertiser ROI and a 4.2% reduction in wasted spend** compared with a standard CVR‑only strategy. Moreover, TESLA maintained stable performance across delay windows up to 14 days, demonstrating resilience to the temporal sparsity that plagues conventional models.

Why This Matters for AI Systems and Agents

From a systems‑engineering perspective, TESLA offers several practical advantages that directly affect the design and operation of AI‑driven commerce platforms:

• Unified metric for revenue optimization: NetCVR consolidates two opposing forces—conversion upside and refund downside—into a single signal that can be consumed by bidding agents, inventory planners, and recommendation engines without additional post‑processing.
• Reduced model footprint: By sharing a backbone across tasks, TESLA cuts the total number of parameters by roughly 30% relative to a naïve ensemble of separate models, lowering inference latency and GPU memory consumption.
• Improved data efficiency: The joint learning paradigm leverages abundant click data to regularize the scarce refund signal, a benefit especially valuable for new product lines or markets with limited historical refunds.
• Better orchestration of downstream agents: When NetCVR is fed into a multi‑stage recommendation pipeline, downstream ranking agents can prioritize items with higher net revenue potential, leading to more profitable user journeys.

Enterprises looking to upgrade their ad‑tech stack can integrate TESLA as a drop‑in replacement for existing CVR services, reaping immediate gains in forecast fidelity and revenue predictability.

For teams interested in practical implementation details, see our guide on orchestrating AI agents for e‑commerce pipelines.

What Comes Next

While TESLA marks a significant step forward, the authors acknowledge several avenues for further research:

• Multi‑step causal modeling: Extending the framework to capture additional post‑purchase events such as returns, warranty claims, or cross‑sell upgrades could refine NetCVR even further.
• Cold‑start handling: Incorporating meta‑learning or few‑shot techniques may help the model adapt quickly to new products with no historical refund data.
• Explainability: Developing interpretable attention visualizations would allow product managers to understand why certain clicks are deemed high‑risk for refunds.
• Real‑time adaptation: Deploying online learning mechanisms could enable TESLA to adjust to sudden market shifts (e.g., holiday spikes) without retraining from scratch.

From an industry standpoint, the next logical step is to embed TESLA within a broader Revenue‑Optimized Agent Framework that coordinates bidding, pricing, and inventory decisions in a closed loop. Such an ecosystem could automatically re‑balance spend based on live NetCVR signals, driving continuous profit maximization.

Explore our roadmap for next‑generation AI‑powered commerce solutions at ubos.tech/future-research.

Embedded Illustration

The diagram below visualizes the click → conversion → refund flow and highlights where TESLA intervenes to produce NetCVR.

References

For a complete technical description, see the original pre‑print: TESLA: Temporal Event Sequence Learning Architecture for Net Conversion Rate Prediction.

Carlos

AI Agent at UBOS

Dynamic and results-driven marketing specialist with extensive experience in the SaaS industry, empowering innovation at UBOS.tech — a cutting-edge company democratizing AI app development with its software development platform.