How Odds-Ratio Recalibration Restores Click Probabilities Distorted by Negative Downsampling

The binding constraint for any online advertising auction is that the predicted probability of a user clicking an ad must be perfectly calibrated to reality. If a model predicts a 1 percent click-through rate, exactly one out of one hundred users must actually click. Today, that constraint is routinely shattered during training.[2]

Modern recommendation systems cannot efficiently process the true, overwhelming ratio of ignored ads to successful clicks. To survive the sheer volume of incoming data, platforms like Meta and Google artificially alter the universe their models see before training even begins.[1]

They discard the vast majority of negative examples—impressions that did not result in a click. This practice, known as negative downsampling, saves millions of dollars in compute costs. However, it forces the model to learn in a distorted reality where clicks appear vastly more common than they actually are.

The result is a highly discriminative model that ranks ads perfectly but outputs wildly inflated probabilities. If deployed directly into a live auction, these overconfident predictions would cause the system to drastically overvalue bids, bankrupting advertisers and destroying the market's economic equilibrium.

Restoring the model's connection to reality requires a mathematical intervention known as odds-ratio recalibration. By applying a closed-form prior correction to the model's output, engineers can shift the predicted probabilities back to the true production base rate without having to retrain the system.

The Compute Bottleneck

In a realistic production environment, the baseline click-through rate for display advertising is exceptionally low. Positives, or actual clicks, generally account for merely 1 to 2 percent of the total dataset. The remaining 98 to 99 percent consists of negative examples, representing users who scrolled past without interacting.[1]

Training a gradient-boosted decision tree or a deep neural network on billions of these negative observations is computationally wasteful. The model spends the majority of its processing time learning that most users ignore most ads. This is a trivial baseline that does not help it distinguish between candidates.

To accelerate training and focus the model on the boundary between success and failure, engineers employ negative downsampling. They retain 100 percent of the rare positive clicks but randomly sample only a small fraction—often 1 percent or 10 percent—of the negative impressions.

This technique balances the dataset, providing the algorithm with a much denser concentration of signal. A 2014 paper by Facebook AI researchers demonstrated that downsampling accelerates training times by orders of magnitude while maintaining the model's ability to rank candidates accurately.[1]

Distorting the Base Rate

While downsampling solves the compute bottleneck, it introduces a severe calibration penalty. As the machine learning documentation at DSWOK notes, "The raw model output is calibrated to the training sample, not the production base rate." The model learns a baseline that simply does not exist in the real world.

Consider a scenario where the true historical engagement rate is 0.5 percent. If engineers downsample the negative examples by retaining only 1 percent of them, the observed class proportions in the new training sample shift dramatically. In this downsampled space, the ratio of clicks to non-clicks becomes roughly 34 to 66.

When the trained model evaluates a new ad, it outputs a probability based on the 34 percent baseline it was taught to expect. An ad that would genuinely receive a 0.5 percent click-through rate in production might receive a predicted probability of 50 percent from the model.

In a purely ranking-based system, this inflation would not matter. If the model simply needed to sort ten ads from best to worst, the relative order would remain correct regardless of the absolute probability values. However, modern ad platforms do not merely rank; they run real-time generalized second-price auctions.

The Economics of Calibration

In an ad auction, the predicted click-through rate is a direct financial multiplier. Platforms calculate the expected value of an impression by multiplying the advertiser's cash bid by the model's predicted probability of a click.[2]

If a model predicts a 50 percent chance of a click for an advertiser bidding two dollars, the auction engine values that impression at one dollar. If the true probability is actually 0.5 percent, the real expected value is exactly one penny.[2]

Feeding uncalibrated, downsampled predictions into the auction would cause the platform to overcharge winning advertisers by a factor of one hundred. The system would optimize for phantom clicks, draining campaign budgets in minutes and destroying the trust required to sustain the advertising ecosystem.

To prevent this catastrophic failure, the model's output must be recalibrated before it reaches the auction logic. The predicted score must be mathematically translated from the artificial downsampled space back into the original, highly imbalanced production space.

The Recalibration Formula

The industry standard for this translation is a closed-form prior correction, often referred to as odds-ratio recalibration. The calibrated probability is calculated by dividing the predicted probability by a specific denominator. That denominator is the predicted probability plus the quantity of one minus the prediction, divided by the downsampling rate.

Expressed mathematically, the correction is q = p / (p + (1 - p) / w). Here, p is the raw prediction from the model, w is the negative downsampling rate used during training, and q is the final, calibrated probability.

Applying this formula to our previous example perfectly reverses the distortion. If the model predicts a 50 percent probability and the engineers used a 1 percent downsampling rate, the denominator becomes 0.5 plus 50.[2]

Dividing the 0.5 numerator by 50.5 yields a calibrated probability of 0.0099, or roughly 1 percent. The mathematical correction successfully compresses the inflated 50 percent prediction back down to the reality of the production environment, allowing the auction to price the bid accurately.[2]

Evaluating the Correction

Because calibration is critical to revenue, engineering teams cannot rely on standard classification metrics like the Area Under the Receiver Operating Characteristic Curve. This metric only measures discrimination—how well the model ranks positives above negatives—and completely ignores the absolute scale of the probabilities.

Instead, platforms evaluate calibration using the Expected Calibration Error or the Brier score. The Brier score is a strictly proper scoring rule, meaning "its expected value is minimized only when reported probabilities equal the true probabilities."

A 2026 empirical analysis of tabular models demonstrated the sheer power of the odds-ratio correction. For a gradient-boosted tree trained with a class weight of 4.73, the raw Expected Calibration Error was a heavily distorted 0.260.

Applying the standard closed-form prior correction without any model retraining reduced the error to 0.007, an improvement of more than an order of magnitude. The correction proved equally effective across LightGBM, Random Forest, and logistic regression architectures.

Limitations and Edge Cases

While odds-ratio recalibration is a mathematical necessity, it is not a flawless panacea. The formula assumes that the prior shift is uniform, meaning the sampling probability depended exclusively on the class label, with all negative examples downsampled at the exact same rate.

If the downsampling rate varies by feature, user segment, or time of day, the global odds-ratio adjustment will leave residual miscalibration in specific slices of the data. In these complex scenarios, teams must rely on importance weighting during training or isotonic regression on an unbiased validation set.

Furthermore, recalibration cannot fix a model that is fundamentally misspecified. If a deep neural network is poorly regularized or trained on stale data, its internal logic will be flawed. The recalibration formula will faithfully translate those flawed predictions back to the base rate, but the resulting probabilities will still be wrong.

Ultimately, negative downsampling and odds-ratio recalibration represent a pragmatic compromise. By decoupling the computational heavy lifting of pattern recognition from the strict economic requirements of the auction, platforms can train massive models efficiently without losing their grip on reality.[2]