I am Bill SerGio, a trailblazer in the world of television marketing and a visionary entrepreneur. I wrote, financed, directed, and produced over 100 groundbreaking infomercials, each featuring famous celebrities showcasing innovative products that I personally developed. Beyond creating these record-breaking campaigns, I also revolutionized how TV time is purchased, employing a neural network and Bayesian optimization algorithms I designed to secure the best airtime deals. My success is unparalleled—22 of my half-hour infomercials alone generated staggering first-year sales ranging from $100 million to $500 million each, cementing my legacy as one of the most successful figures in the history of direct-response advertising.

Some basics about buying half-hours of television time is that it is actually an auction under federal law that defines the commercial sale of television time as a "Negotiated Price Product."

Negotiated Price Product refers to the federally mandated process by which television and radio advertising time is sold. Under U.S. federal law, broadcasters are prohibited from selling advertising time at fixed rates through a rate card system. Instead, prices must be determined through negotiation between the buyer and seller. This ensures that rates are flexible and subject to market dynamics, typically reflecting factors such as demand, time slots, audience size, and competitive bidding.

Most people will call an ad agency, and the agency will make you pay the "rate card" prices for television time. It is a felony to issue a "rate card" with fixed prices for television time, so never buy any television time from any ad agency.

Other people will call a television station and ask, "How much?" for a half-hour of television time. And the salesperson will ask, "What is your budget?" Whatever your budget is, the salesperson will tell you that is what a half-hour costs.

A sophisticated buyer of television time will simply call a television station and tell the salesperson to email a list of their avails and fire sales without including any prices because the buyer will tell them what they are willing to offer for those time slots.

The way I buy television time is to bid on given half-hour time slots. A typical half-hour of television time on the largest broadcast television stations will range from $20 to $200. Are you surprised? Half-hours are cheap because it will cost a television station a lot of money to buy or produce a half-hour show to fill that slot. Selling that half-hour to a buyer who already has a half-hour show produced saves the television station that cost.

The math analysis problem I wanted to solve was how we can determine the maximum amount to bid on any half-hour so that we can achieve a certain "pull ratio" and guarantee a net profit over the cost of the media.

Pull Ratio: The pull ratio is a performance metric used in advertising, defined as the gross sales generated divided by the cost of the media used to promote the product or service. For example, a pull ratio of 2.0 means that for every dollar spent on media, two dollars in gross sales were generated. It is commonly used to assess the effectiveness and profitability of advertising campaigns.

How did I write a neural network to accomplish this? I will explain how I did it in today's terminology so you will understand why you can actually guarantee a net profit by buying half-hours of television time.

The key to accomplishing this is the FACT that the pull ratio of an infomercial is CONSTANT. If you test an infomercial on a "test station" and it pulls, say, 2 to 1, then it will always have that same pull ratio until one day the pull ratio will be 0, and the show has run its course. This fact makes it possible to use a neural network with Bayesian optimization to predict with great accuracy across thousands of media buys how to allocate your media budget to guarantee an overall net profit across your media buys.

A neural network, combined with Bayesian Optimization, can effectively distribute your $50,000 budget across television stations to maximize net profits. This approach models the intricate relationships between input features—such as pull ratios, costs, and station-specific constraints—and outputs, such as profit. The fixed pull ratio of each infomercial on each station, once aired at least once, provides a reliable foundation for precise predictions. This enables you to make data-driven decisions that outperform simpler heuristics.

Why Neural Networks and Bayesian Optimization Work Together

• Reliable Pull Ratios: Each station has a fixed pull ratio for an infomercial, determined after the first airing. Neural networks can use these fixed ratios to learn station-specific patterns, while Bayesian Optimization explores budget allocation strategies to maximize profit.
• Historical Data as a Foundation: With data from 1,000 stations, including pull ratios, costs, and profits, a neural network identifies the factors that drive profitability. Bayesian Optimization refines this process by systematically searching for the most effective allocation strategies.
• Constraint Modeling: Constraints like one airing per week and resting periods introduce temporal dynamics. Neural networks, especially recurrent architectures like LSTMs, model these time-based patterns. Bayesian Optimization ensures compliance with constraints while optimizing outcomes.
• Profit Maximization Beyond Pull Ratios: While high pull ratios are advantageous, they are not the sole determinant of profit. Factors such as station fatigue, diminishing returns, and cost must also be considered. This is where the neural network and Bayesian Optimization excel, delivering balanced, high-performing strategies.

Approach to Implementation with ML.NET

1. Data Preparation:
   • Inputs:
     • Station data: Pull ratio, cost per half-hour, historical performance.
     • Infomercial data: Product type, previous station performance.
     • Temporal data: Days since last airing, cumulative airings.
   • Outputs:
     • Net profit (target variable).
2. Feature Engineering:
   • Normalize pull ratios and costs for better model convergence.
   • Encode categorical variables (e.g., station IDs, infomercial types).
   • Include time-series features like cumulative airings and intervals between airings.
3. Optimization Layer: Use Bayesian Optimization to iteratively allocate the budget while maximizing profit predictions from the neural network.

Code Implementation

Neural Network Training in ML.NET

using Microsoft.ML;
using Microsoft.ML.Data;

public class StationData
{
    [LoadColumn(0)] public string StationID;
    [LoadColumn(1)] public float PullRatio;
    [LoadColumn(2)] public float CostPerHalfHour;
    [LoadColumn(3)] public float DaysSinceLastAiring;
    [LoadColumn(4)] public int TotalAirings;
    [LoadColumn(5)] public float Profit; // Target variable
}

public class ProfitPrediction
{
    [ColumnName("Score")] public float PredictedProfit;
}

var mlContext = new MLContext();

// Load and preprocess data
var dataPath = "station_data.csv";
var data = mlContext.Data.LoadFromTextFile(dataPath, separatorChar: ',');

// Split data for training and testing
var split = mlContext.Data.TrainTestSplit(data, testFraction: 0.2);

// Define and train the model
var pipeline = mlContext.Transforms.Categorical.OneHotEncoding("StationID")
    .Append(mlContext.Transforms.Concatenate("Features", "PullRatio", "CostPerHalfHour", "DaysSinceLastAiring", "TotalAirings"))
    .Append(mlContext.Regression.Trainers.FastTree());

var model = pipeline.Fit(split.TrainSet);

// Evaluate the model
var predictions = model.Transform(split.TestSet);
var metrics = mlContext.Regression.Evaluate(predictions, "Profit");

Console.WriteLine($"R^2: {metrics.RSquared}, RMSE: {metrics.RootMeanSquaredError}");

Integrating Bayesian Optimization

// Example using Accord.NET for Bayesian Optimization
using Accord.MachineLearning.Optimization;

Func profitFunction = (parameters) =>
{
    var pullRatio = (float)parameters[0];
    var costPerHalfHour = (float)parameters[1];

    var predictionEngine = mlContext.Model.CreatePredictionEngine(model);

    var stationData = new StationData
    {
        StationID = "SampleStation",
        PullRatio = pullRatio,
        CostPerHalfHour = costPerHalfHour,
        DaysSinceLastAiring = 7,
        TotalAirings = 1
    };

    var prediction = predictionEngine.Predict(stationData);
    return prediction.PredictedProfit; // Objective to maximize
};

var optimizer = new BayesianOptimization(profitFunction, 
    lowerBounds: new[] { 1.0, 50.0 }, 
    upperBounds: new[] { 10.0, 500.0 });

var result = optimizer.Maximize();
Console.WriteLine($"Optimal PullRatio: {result.Solution[0]}, Optimal CostPerHalfHour: {result.Solution[1]}");

Summary

Leveraging a neural network's predictive capabilities alongside Bayesian Optimization ensures that every dollar of your $50,000 budget is allocated to maximize net profits across all media buys. The fixed pull ratio on each station for a given infomercial, determined after its initial airing, provides a solid foundation for accurate forecasting and strategic optimization when combined with an overriding rules engine.

An AI Rules Engine using the Census database guarantees that any given infomercial selling a product is precisely targeted to the matching customer profile. I only bought high-power broadcast stations with 50,000-watt transmitters, which provide a 200-mile broadcast radius.

The AI Rules Engine identifies all ZIP codes within a 200-mile radius of the latitude and longitude of each station's transmitter. It retrieves Census demographics for those ZIP codes, enabling precise targeting. The next step involves matching the social profile of the host of the infomercial and the product's target demographics with the Census data for potential broadcast stations. Analysis revealed that men and women purchase equally from any infomercial, and the two critical factors to align were race and income.

As a quick note, we can also match the Census demographics to newspapers and magazines and the Internet. But more that in a separate article with M.NET source code I will publish.