Inside the Black Box: The Statistical Modelling Behind Automated Bid Management Tools

Articles No Comments

While today’s PPC management systems are providing increasingly comprehensive toolsets, automated bid management remains the primary reason most advertisers and agencies rely on such systems. The prospect of manually optimizing bids across thousands or possibly millions of keywords without such automation is a daunting if not impossible task to perform effectively.   And while most of the leading bid management system vendors provide relatively advanced automated bid management capabilities, the campaign managers that trust in them are often left in the dark regarding what’s actually going on behind the interface.

This whitepaper intends to provide insight into many of the challenges and resulting processes that are employed by leading bid management tools.

Algorithmic vs. Rules-Based Bidding

Before delving into the concepts behind automated bidding, it is important to differentiate an algorithmic approach from a rules-based approach. Both are methods to automate bid management, but we will only be discussing the algorithmic approach in this paper for two primary reasons. First, rules-based automation is highly transparent since the users typically set the rules up on their own. Therefore, there is little mystery for this approach. Secondly, rules-based automation is far more limited than algorithmic automation. Algorithmic bidding is able to consider more data elements and process more advanced computations. Rules-based bidding cannot perform something even as elementary as a correlation coefficient to determine the extent to which different metrics are associated. Therefore, only the mysterious algorithmic approach will be discussed here.

The Perfect Bid

If we had the ability to see into the future, there are only two primary metrics we would need to identify to set the perfect bid for every keyword to maximize volume within a target CPA – conversion rate and bid gap (average order price would also be necessary for ecommerce advertisers).

To illustrate this, let’s consider an example of a lead gen advertiser that wants to drive as much conversion volume possible while maintaining a $10 CPA. If we could accurately predict tomorrow’s conversion rate and bid gap for a particular keyword, we could calculate the exact bid necessary to achieve the highest position at the target $10 CPA.

In my crystal ball, I have seen that tomorrow’s performance for Keyword X is the following. . .

Conversion rate = 5%
Bid gap = $0.25

Knowing that my target CPA is $10, I can easily calculate the CPC I need to pay by multiplying by the conversion rate.

$10 Target CPA * 5% conversion rate = $0.50 CPC

Now I just need to add in the bid gap to find the amount I need to bid.

$0.50 CPC + $0.25 bid gap = $0.75 CPC Bid

This is the perfect bid. Once completed for all keywords, you can expect money to rain from the sky tomorrow.

The Rub

Unfortunately, we can’t see into the future. Instead, we must predict future conversion rates and bid gaps by analyzing historical performance data. This is where the power of advanced bidding algorithms becomes imperative.

Situations All Bidding Algorithms Must Address

In many situations, bidding algorithms can be highly accurate at predicting future conversion rates at the keyword level. But there are many more situations where accurate predictions become extremely difficult and require advanced statistical analysis. While the strategies for dealing with these situations vary considerably between one bid management system to another, all systems must contend with them one way or another.

While not an exhaustive list, below are some of the primary issues bidding algorithms must address to make accurate predictions of future performance.

  • Volume of historical data to consider
    • Statistical significance required for desired confidence levels
    • Methodology for predicting performance of keywords not meeting statistical significance requirements
  • Time period of historical data to consider
  • Variance of key metrics over time
  • Aggressiveness of bid adjustments when considering probability and margins of error

Historical Data Set Selection

How much historical data should be considered for each keyword? This question must be addressed from both a volume and time perspective. Let’s first delve into the volume side of this question.

The methodology for managing low volume keywords is a major concern within the design of bidding algorithms. If a keyword has received two clicks and one conversion over the past 30 days, we certainly wouldn’t want to predict tomorrow’s conversion rate at 50%. The algorithm must consider the click and conversion volume of every keyword to know the level of accuracy we can expect from our prediction. A keyword with 1,000 clicks over the last 7 days will likely result in a highly accurate prediction. Historical data for a keyword with only 10 clicks will certainly provide a poor prediction. And even more troublesome, how will conversion rates be predicted for keywords that haven’t received any conversions? Predicting a 0% conversion rate will result in a keyword bid being set to $0.00. But if the keyword has only received 49 clicks, a conversion on the next click would make the average conversion rate jump from 0% to 2% leading to a dramatic difference in bidding. This is essentially a question of determining 1) statistical significance of click volume to achieve an acceptable and predictable margin of error and 2) bidding logic to consider lower levels of significance and how to appropriately adjust bidding strategies to limit losses resulting from uncertain predictions.

The most common strategy for dealing with low volume keywords is to consider the conversion rate of groups of keywords when individual keywords do not have sufficient data to make accurate predictions on their own. This must be a two-stage process, first defining the weighting ratio between individual keywords and groups. The second stage is determining which group of keywords will be considered to support the bid decision of an individual keyword. A good bidding algorithm will have a variable weighting calculation based on the level of volume for each keyword. In this way, the algorithm will weight individual performance for a keyword with 100 clicks heavier than for a keyword with 10 clicks.

Strategies for addressing the second part of the process are less obvious. One method of determining which groups of keywords will be considered to predict long-tail performance is to group according to the existing campaign structure. In other words, the algorithm will first consider performance data for the entire ad group and subsequently performance of the entire campaign or account. While this is a workable method, it assumes that keywords within a particular ad group or campaign perform similarly, which very often is not the case. A better method is for the algorithm to first identify commonalities in keywords that tend to result in similar performance and group based on these commonalities. For example, keywords including the term “buy” may consistently convert better than those without regardless of which ad group they belong to. This is an advanced method that requires intelligent language analysis in addition to performance analysis.

Time Period of Historical Data

An additional consideration when selecting a historical data set is the time period the data covers. We may be able to obtain a high volume of click data for a specific keyword if we span a two year time period.   But this seemingly significant volume may not be a good performance predictor due to changes in the marketplace, seasonality, and general user behavior over time. A strategy must be put in place to ensure the algorithm includes a statistically significant amount of data, but not so much history that changes in the market cannot be identified. One day will likely not provide enough data for most keywords. A year of historical data won’t be able to factor in recent market changes.

Variance

The daily variance of key metrics is also a major consideration in determining the accuracy of a prediction. If a keyword only varies one to two points above and below the average conversion rate for the last three months, we can be highly confident that tomorrow’s conversion rate will also be within one to two points of the average. But for low volume, low converting keywords, conversion rates may be jumping from 0% one day to 25% the next. This reduces our confidence in the historical data’s ability to accurately predict tomorrow’s conversion rate.

Bidding Aggressiveness

Once all of the above situations are considered and factored into conversion rate and bid gap predictions, an intelligent bidding algorithm will want to ensure further accuracy by pacing bid changes towards the optimal bid calculated from historical data. In other words, a dramatic increase in conversion rates might result in a bid change recommendation moving from $1.00 to $1.50. However, the dramatic improvement may be short-lived due to market pressures (competitor dropping out of auctions), outside influences (Oprah endorses the product on her show), or seasonal patterns (higher conversion rates prior to a holiday). For this reason, the bidding algorithm may want to approach the estimated optimal bid rather than changing as abruptly as performance data.

Methods for pacing the speed at which algorithms approach predicted target bids vary widely from system to system. Furthermore, factors that influence bidding aggressiveness may be driven by campaign management goals instead of purely data-driven optimization. For this reason, purchasers of automated bidding systems must consider flexibility and customization capabilities when evaluating such systems.

Summary

This paper has addressed the fundamental approach automated bidding algorithms use to optimize keyword bids towards an ROI metric. A more detailed discussion of the process would require us to delve into statistical formulas and database programming topics. However, our goal was not understand the intricacies of designing bidding algorithms, but merely to take the mystery out of what’s going on behind the scenes. We hope that this paper will help frontline campaign managers understand how bid adjustment decisions are being made as well as communicate to clients the importance of using such tools.

Author

Leave a comment

Back to Top