Hedging Between Budget Pacing Controllers, Checked Across Open Auction Benchmarks

Advertisers who buy ads through real-time auctions give an automated bidder a daily budget. That bidder needs a pacing rule, meaning some way to avoid spending the whole budget on expensive morning traffic and having nothing left for the cheaper, higher-converting evening hours. Many pacing rules already exist, and the public benchmark BAT: Benchmark for Auto-bidding Task shows that none of them wins everywhere. Mystique spends most evenly. Traffic-Aware PID (TA-PID) and the Adaptive Linear Model (ALM) get the most clicks. Model-Predictive PID (M-PID) best respects cost-per-click caps in first-price auctions. Each of these controllers outputs the same kind of quantity, a pacing multiplier $\alpha$ that scales every bid. That makes the obvious question cheap to ask. Can a small online learner, which reweights the controllers' multipliers as the day unfolds, beat every individual controller and also beat simply picking the best controller for each campaign?

We propose that learner and evaluate it under a reproducibility discipline the area currently lacks. The combiner hedges across controllers in "multiplier space" at the pacing-update cadence, which is minutes, not per bid, so it adds almost nothing to bid latency. The ensemble is judged on more than average performance. We also measure whether its advantage, and the ranking of methods, holds across two independently built public benchmarks, BAT and AuctionNet, and across random seeds. In a field where the strongest systems are validated on private logs, a method whose lead disappears when the benchmark changes has not shown much.

Core Claim

The controllers' different strengths in BAT reflect regime dependence, not a fixed trade-off. Which controller is best changes with the campaign, the hour, the traffic level (BAT documents a 30x peak-to-off-peak swing) and the auction mechanism. If that is true, an online combiner with a no-regret guarantee against the best controller in hindsight should match or beat the per-campaign best controller without knowing in advance which one that is. It should also be more stable across benchmarks than any single controller.

If the strengths instead reflect a fixed trade-off, the combiner will only land somewhere on the existing trade-off curve, and a single controller retuned for the same objective will do as well. Distinguishing these two outcomes is the scientific point of the project.

This is distinct from everything in the file. Feedback Control of Real-Time Display Advertising and Smart Pacing for Effective Online Ad Campaign Optimization each design one controller. An Adaptable Budget Planner for Enhancing Budget-Constrained Auto-Bidding in Online Advertising (ABPlanner) learns a stage-level budget plan but hands execution to a single fixed bidder, either PID or a linear-programming solver. Auto-Bidding in Real-Time Auctions via Oracle Imitation Learning (OIL) trains one student network to imitate one hindsight oracle. Neither BAT nor AuctionNet (AuctionNet: A Novel Benchmark for Decision-Making in Large-Scale Games, companion entry) evaluates any combination of controllers.

The closest methodological ancestor sits outside advertising. Adaptive expert fusion model for online wind power prediction reweights online forecasters. We borrow that idea but apply it to controllers whose outputs feed back into the state they observe. That feedback creates a problem forecasting does not have, addressed in the next section.

Research Program

The pool of experts contains the five BAT controllers: ALM, TA-PID, M-PID, Mystique, and BROI (a controller that paces both budget and return on investment). It also contains an online linear-programming pacer and, where it can be run, a trained OIL policy. Every expert is converted to emit a multiplier $\alpha_{k,t}$ for expert $k$ at update step $t$.

The combined multiplier is a weighted geometric mean, which keeps $\alpha_t$ inside the range the experts span and treats halving and doubling symmetrically:

$$\log \alpha_t = \sum_{k} w_{k,t}\,\log \alpha_{k,t}.$$

Here $w_{k,t}$ are nonnegative weights summing to one. They are updated by exponential weighting with a small "fixed-share" mixing step, which lets weight move back to an expert that becomes good again after a regime change. The update is $w_{k,t+1} \propto w_{k,t}\exp(-\eta\,\ell_{k,t})$, followed by mixing a fraction $\gamma$ of mass uniformly across experts. The learning rate is $\eta$.

The loss each expert is charged comes from replaying the last window under that expert's multiplier:

$$\ell_{k,t} = -\,\text{clicks}_{k,t} + \lambda_t\,\lvert \text{spend}_{k,t} - \text{target}_t \rvert + \mu_t\,\max(0,\ \text{spend}_{k,t} - c\cdot\text{clicks}_{k,t}).$$

In this loss, clicks and spend are what the campaign would have obtained in the window under multiplier $\alpha_{k,t}$. The term $\text{target}_t$ is the budget-trajectory target for that window, and $c$ is the campaign's cost-per-click cap. The weights $\lambda_t$ and $\mu_t$ are dual prices, raised when realized pacing error or cap violation accumulates. They make the scalarization adaptive rather than hand-set.

Two technical issues need care.

The first is counterfactual evaluation. Charging every expert requires knowing what its multiplier would have won. In BAT log replay the competing prices are logged, so this is exact under the standard assumption that one campaign does not move the market. OIL makes the same exogenous-price assumption. In AuctionNet's 48-agent environment that assumption breaks, which is precisely why we test there as well.

The second is controller state. A PID controller whose output is not applied will accumulate a stale integral term. Every expert is therefore driven by the realized spend trajectory, with its integral reset to that trajectory (standard anti-windup), so each expert gives its recommendation for the actual current state.

Standard results for exponential weighting give regret growing as the square root of the number of update steps against the best fixed expert per campaign. That is the formal sense in which the combiner should match per-campaign selection without knowing the right expert beforehand. Fixed-share extends the guarantee to the best switching sequence of experts.

The reproducibility lens is part of the evaluation design. The cautionary example in the file is Deep Ensemble Learning for the Automatic Detection of Pneumoconiosis in Coal Worker's Chest X-ray Radiography. Its ensemble gain, 91.50% versus 90.20% accuracy for the best single network, was measured on 153 private radiographs and cannot be checked. Pacing papers face the same risk: ABPlanner's main evidence comes from 20,000 private advertisers and a six-day live test.

We will therefore report seed-level variance for every number, and give every main result on BAT first-price, BAT VCG, and AuctionNet's GSP auctions (generalized second-price, the multi-slot format AuctionNet uses). We will also report rank agreement between benchmarks. One further probe targets anonymization. BAT shifts timestamps for confidentiality, while date-shifting in MIMIC-III, a freely accessible critical care database preserves intervals but not true clock position. Controllers that rely on a learned hour-of-day traffic curve, such as TA-PID, may therefore be partly fitting an artifact. Applying additional controlled time-of-day shifts to the BAT traffic curve tells us how much each method depends on that curve, and whether the combiner absorbs the dependence by moving weight elsewhere.

Baselines

The do-nothing reference is the unpaced bidder, with $\alpha=1$ until the budget runs out. It shows how much any pacing buys on each benchmark, and the 30x traffic swing in BAT suggests the answer is a lot. Alongside it we run each BAT controller alone, with the hyperparameters BAT publishes.

The strongest existing method on AuctionNet is OIL. The file reports that OIL improves on the best of PPO, IQL and online linear programming by 21.2% relative on AuctionNet-dense and 17.8% on AuctionNet-sparse, with about 99% budget use. We rerun OIL from its description on the same AuctionNet splits and seeds, rather than copying those numbers, and include it both as a baseline and as an expert in the pool. Beating OIL on AuctionNet means a higher campaign score than OIL alone at matched budget utilization, with a seed-level confidence interval that excludes zero. For BAT, no numeric controller results were found on Lacuna. BAT reports only rankings: Mystique has the lowest RMSE, TA-PID and ALM get the most clicks, and M-PID best satisfies CPC caps in first-price auctions. We therefore regenerate all BAT numbers ourselves under one shared protocol.

The baseline most likely to match us is per-campaign best-controller selection, run in two forms. The oracle form chooses the controller that did best on that campaign in hindsight. It is an upper bound for any selection method, and the combiner can exceed it only if switching within a day helps. The learned form is a classifier that predicts the best controller from campaign attributes, budget, and auction mechanism, trained on held-out campaigns. Beating the oracle selector shows that within-day switching matters. Beating the learned selector shows that online hedging is worth more than offline algorithm selection.

The single-controller rival receives the same objective. TA-PID and M-PID are retuned against the loss above, with a hyperparameter search budget equal to what the combiner's $\eta$, $\gamma$ and dual step sizes receive.

The ablations isolate the two key ingredients. Uniform geometric averaging of multipliers, with no learning, tests whether adaptation matters at all. Hedging with exact replay losses replaced by losses computed only from the realized trajectory tests whether counterfactual replay matters.

Why not just…?

Why not just pick the best controller per campaign? This is the real threat. If controller strengths are a stable property of the campaign rather than of the hour, the learned selector will match the combiner at lower complexity, and we will say so. The decisive comparison is combiner versus oracle per-campaign selection on BAT. If the combiner never exceeds the oracle, within-day regime switching is not real, and the paper's recommendation becomes "learn a selector."

Why not just retune one PID controller for the combined objective? Retuning moves a single controller along its own trade-off curve. The BAT rankings show that different controller designs, not different settings of one design, occupy the best regions for evenness, clicks, and cost caps. Whether retuning closes the gap is answered directly by the equal-budget retuned TA-PID and M-PID baselines, reported as trade-off curves of pacing error against clicks rather than as single points.

Why not just use one learned policy such as OIL or ABPlanner? OIL is strong on AuctionNet. However, it imitates a single oracle that assumes fixed competitor prices, and it has only been shown on GSP auctions. ABPlanner's strongest evidence is private and its execution layer is a single solver. Including OIL as an expert means the combiner cannot do worse than OIL by more than its regret, so the test is whether it adds anything. If OIL alone matches the combiner on AuctionNet but OIL transfers poorly to BAT first-price auctions, that cross-benchmark gap is itself the finding.

Why not just average the bids? Arithmetic averaging of bids ignores how first-price and GSP auctions clear and can break budget and cost-per-acquisition constraints. Averaging multipliers is the sensible naive version, and it is our uniform-weight ablation.

Proposed Experiments

On BAT first-price (about 9,000 campaigns) and BAT VCG (about 2,500 campaigns), we run all methods in log replay with campaigns split into tuning and test sets. We report budget-trajectory RMSE, total clicks, CPC-cap violation rate, and the combined score, each with seed-level intervals.

On AuctionNet we run the same methods in the multi-agent environment. Here the exogenous-price assumption behind counterfactual replay fails, so this is a stress test. A large drop relative to BAT would locate the method's limit. We also report AuctionNet results separately on high-value opportunities, because the AuctionNet authors note their diffusion generator reproduces rare VIP users poorly. Gains concentrated there would be suspect.

Four checks decide the project. First, does the combiner beat oracle per-campaign selection on BAT, which is the test of within-day switching? Second, does it beat the equal-budget retuned single controllers on the clicks-versus-RMSE trade-off curve, which is the test of regime dependence against a fixed trade-off? Third, does it beat OIL on AuctionNet when OIL is in its pool? Fourth, is its rank across BAT first-price, BAT VCG and AuctionNet more stable than that of any single controller, measured by rank correlation across benchmarks and by variance across seeds?

A negative answer on the first two checks is publishable and useful. It would tell practitioners that pacing controllers differ by trade-off, not by regime, and that the right tool is a well-chosen single controller. We release code, seeds, and the shared evaluation protocol for all three benchmark settings.