## What is direct mail marketing?

Direct mail marketing is the act of a brand engaging with its customers and prospects through physical mail. It can come in various forms, from postcards, to letters, to branded water bottles, and everything in between. The most effective direct mail campaigns are highly personalized to resonate with a brand’s customers. Before email and digital engagement became the norm, direct mail was one of the best ways to reach your customers. While digital has since pushed direct mail marketing to the side, it’s had a resurgence in recent years because, in the age of digital media, personalized direct mail _stands out_–assuming you’ve done your due diligence and personalized it appropriately to match your customers’ wants and needs. Take [PostcardMania](https://www.postcardmania.com/blog/direct-mail-statistics-2023) for example, who has seen their revenue grow 17.56% each year post-pandemic in comparison to 4.67% pre-pandemic.

## Benefits of personalized direct mail

Personalized direct mail has a variety of benefits over digital to prove that it’s a great supplement to your engagement strategy. To start, the response rate of direct mail marketing is 9%, in [comparison](https://www.postgrid.com/direct-mail-statistics/) to the 1% of email, paid search, and social media. Additionally, it has a 29% ROI compared to 23% for paid search and 16% for online display. The benefits of personalized direct mail stem from two key features:

The first is that personalized direct mail can be tailored to a greater extent than its alternatives. Say that you know a group of your customers are avid runners that regularly buy running gear from your store. A digitally-focused brand might launch a campaign between major purchases for that group of people, where they want to offer them a free water bottle for being a loyal customer. An email offering the freebie will look a lot like other emails they receive, so it’ll likely be discarded, even if it’s genuine. With a brand savvy in personalized direct mail, the water bottle could be at their doorstep–delivered in a box personalized with running-focused creative–then in their hands as it becomes their daily drinker that they freely advertise for you.

The second is that, in the digital age, people are used to receiving all manner of marketing emails, to the point that it’s almost second nature to mark something as spam, or delete it. Personalized direct mail is more distinctive, and is much more likely to be seen by your customers. After all, 77% of people sort their physical mail immediately [according](https://www.pb.com/docs/US/pdf/Microsite/Nonprofit/ed_np_getyourmailopened_05MailMoment.pdf) to the USPS. With a chance that high for your mail to be seen, making use of personalized direct mail that resonates with your customers is the key to standing out.

## How can direct mail be personalized?

Personalized direct mail can be accomplished through several different means that utilize your first-party data, such as merge tags and creative variations. With machine learning and third-party consumer data on top of your first-party data, you can accomplish it more effectively, and at scale.

### Leverage data and machine learning for insights to guide personalization

To start, your customer data needs basic identifiers, like name and address, so that you know where to send your personalized direct mail. On top of that, you need to leverage all of your first-party engagement data, like products purchased, abandoned carts, subscriptions, and so on, to ensure any offers you make are relevant. You wouldn’t want to find a postcard with a discount in your mailbox for cat food when you don’t own a cat, would you?

Regardless of how strong your first-party data is, it provides limited insight into who your customers are beyond their interactions with your brand. With [rich, third-party consumer data](https://faraday.ai/features/data) that includes demographics, psychographics, lifestyle, financials, and more, you can take your segmentation to the next level, enabling you to customize your copy & creative so that it resonates.

In addition to leveraging both first-and-third-party data, using machine learning to identify distinct groups–personas–allows you to create personalized direct mail at scale, with each group can be targeted with its own resonant, relevant engagement.

### Use insights to inform your personalization

A uniquely powerful way to personalize direct mail is to make use of [merge tags](https://heypoplar.com/articles/personalized-direct-mail). With merge tags, you can personalize on an individualized basis so that any data you have that’s relevant to your customers can be utilized. This way, beneath the persona-level segmentation, your personalized direct mail is dynamic enough to include datapoints like, say, someone’s name, subscription level, or the city they live in. On top of that, if you have physical locations, merge tags can be used to point customers to your nearest storefront.

![Poplar merge tag postcard](/images/blog/direct-mail-merge-tag-poplar.png)

_Image via [Poplar](https://heypoplar.com/articles/personalized-direct-mail)._

Personas also enable distinct creative variations that are relevant to each persona. For your Exercise Guru Elayne persona, you could include imagery of women running alongside a dog and while pushing a stroller, because the majority of the persona are dog owners and have children in the household.

Take your personalized direct mail even further by adding customized product offers and discount codes that are relevant to each persona. Maybe, utilizing your customer data, you know that a specific group of your customers have already purchased a pair of running shoes. You know they won’t need another pair for a while, so a promo for that will bounce off. An offer for a pair of running socks or shorts, on the other hand, is more likely to catch their attention.

![Postcard with personalized discount](/images/blog/direct-mail-discount.png)

_Image via [TheMailShark](https://www.themailshark.com/product-samples/postcards/)._

## Pillars of a personalized direct mail strategy

There are three pillars to a successful personalized direct mail strategy.

- **Insights**: Identify distinct customer segments and their related insights by generating [AI-powered personas](https://faraday.ai/predict/category/persona_assignments). Through these insights, you have clear indicators for how to engage your customers and leads in ways that are highly resonant with their lifestyles.
- **Customization**: Determine how you’ll customize direct mailers based on your insights, data, and promotions. The most impactful personalized direct mail campaigns will include a mix of products, imagery, discounts, and merge tags, ensuring you’re including the right content. Brands that offer personalized direct mail services, like [Poplar](https://heypoplar.com/), provide a range of these types of services.
- **Targeting**: The cost of direct mail can add up–especially if you’re not targeting the best fits. Predict for likelihood to [buy](https://faraday.ai/predict/category/likely_buyers), [convert](https://faraday.ai/predict/category/lead_scores), [repeat purchase](https://faraday.ai/predict/category/repeat_purchasers), [churn](https://faraday.ai/predict/category/churn_scores), or any other relevant datapoint so that you’re spending your time and money where they matter most.

## How to implement a personalized direct mail strategy with Faraday

With a customer prediction platform like Faraday, you can identify personas to gain rich insights and predict behaviors so that you can deliver resonant, personalized direct mail to likely-to-act audiences. All it takes is some historical data about your customers. Here’s how it works.

### Connect data and define customers

To start, you plug in your data by mapping your data fields to Faraday properties.

[![Direct mail dataset](/images/blog/direct-mail-dataset.png)](/images/blog/direct-mail-dataset.png)

Then, in cohorts, you’ll utilize both your customer data and Faraday’s rich, built-in consumer data to build the exact [cohort](https://faraday.ai/docs/cohorts/customers_from_hosted_csv)–or interesting group of people–you want. In Faraday, you’ll create a cohort for your customers so that you can find more people like them, and you’ll also build a cohort of people with mailable addresses, so that you’re also targeting unknown audiences. You'll do this by choosing the following traits when building a cohort:

- **Record quality code**: one of: mailable

[![Faraday direct mail cohort traits](/images/blog/direct_mail_guide/direct-mail-guide-cohort-traits.png)](/images/blog/direct_mail_guide/direct-mail-guide-cohort-traits.png)

### Predict personas to reveal rich insights

While you likely have customer data that indicates, in some ways, who your customers might be, predicting personas tells you who they really are, and what makes them tick. Dive into trait breakdowns like age, gender, household income, hobbies, and more–each one guiding you toward engaging with the right messaging.

[![Direct mail persona set](/images/blog/direct-mail-personas.png)](/images/blog/direct-mail-personas.png)

### Predict propensity to buy to target best fits

Next, you’ll build your predictive outcome by defining your attainment cohort as your customers, as that’s your prediction’s finish line–you want more people like your customers. In this example, we’re defining the propensity to buy, or the likelihood that anyone in the U.S. is going to become a customer by making a purchase, but you can predict for virtually any action that someone is likely to take, such as a conversion, repeat purchase, or churn.

[![Direct mail outcome](/images/blog/direct-mail-outcome.png)](/images/blog/direct-mail-outcome.png)

With your prediction defined, create a pipeline to make them actionable by choosing your **likely buyers outcome**, your **persona set**, and your **mailable cohort**.

[![Direct mail pipeline](/images/blog/direct-mail-pipeline.png)](/images/blog/direct-mail-pipeline.png)

Faraday’s machine learning algorithms then predict how likely people are throughout the U.S. to buy your product, while also revealing which people belong to what persona on an individual basis.

### Deploy personalized direct mail predictions to your stack

Likely buyer scores revealing your best fits across the country? Check. AI-generated personas providing clear guidance for personalization? Check. With these insights in-hand, deploy them anywhere in your stack you need them–for example, to a direct mail provider like our friends at Poplar–then, design your copy & creative so that it caters to the personas you’re engaging with.

**Ready to try it out? [Create a free account](https://faraday.ai/signup).**

Learn the benefits of personalized direct mail marketing, how to personalize your direct mail campaigns, and how to use predictive analytics to maximize impact.

Personalized direct mail: How can direct mail be personalized at scale?

In the world of machine learning, we often encounter two important concepts: the raw propensity score and probability. Though they may sound similar, they represent distinct aspects of a machine learning model's predictions.

## Example of propensity scores vs predicted probabilities

Imagine you have a model that predicts whether it will rain tomorrow. The raw propensity score is like the model's initial hunch or gut feeling about the outcome. It's a numeric value that the model calculates based on various factors like historical weather data, humidity, temperature, and more. The [propensity](https://faraday.ai/blog/propensity-models-predict-customer-behavior) score indicates how likely the model thinks it will rain or not. However, the propensity score itself doesn't give us a clear understanding of how confident the model is in its prediction.

This is where [probability](/blog/faraday-probability-scores) comes into play. The predicted probability aims at providing an answer to the question “how likely is the prediction?” It represents the likelihood of a particular outcome occurring. For example, a score of 0.75 for rain means the model is fairly confident it will rain tomorrow. However, a probability of 0.75 indicates that there is a 75% chance that it will rain tomorrow.

In summary, the raw propensity score is the model's initial guess, while probability quantifies the model's confidence in that guess. Understanding this difference can help us interpret machine learning [predictions](/features/predictions) and make better-informed decisions based on the model's level of certainty.

## Understanding model calibration

The model calibration process refers to the alignment between a model's predicted probabilities and the actual likelihood of events occurring. A good calibration ensures that the predicted probabilities are close to the true theoretical probabilities. Not all machine learning models are perfectly calibrated after training. Some models might be overconfident, providing probabilities that are too extreme, while others might be underconfident, consistently underestimating or overestimating probabilities.

### Importance of model calibration

Calibration is crucial because it impacts decision-making. If a model is well-calibrated, we can trust its probability estimates and use them to make informed choices, such as whether to take an umbrella based on the probability of rain.

Calibrating a model involves post-processing its probability scores to bring them in line with the true probabilities. Various calibration techniques can be employed, such as Platt scaling or isotonic regression, which adjust the raw propensity scores to produce well-calibrated probabilities. These techniques are typically applied during model validation and fine-tuning to ensure that the model's predictions are more accurate and reliable.

The calibration process ensures that when the model predicts a probability, it reflects the true likelihood of the event occurring. For example, if a model predicts a 70% probability of rain, it should be accurate in approximately 70% of the cases where it makes that prediction.

In summary, model calibration in machine learning is all about making sure a model's predicted probabilities accurately reflect the real-world likelihood of events happening. A well-calibrated model is more reliable and can lead to more effective decision-making.

## Benefits of using predicted probabilities

Using predicted probabilities instead of raw propensity scores offers several significant benefits that enhance the interpretability, decision-making, and overall performance of the models:

### Interpretability

Predicted probabilities provide a more intuitive and understandable representation of a model's output. Raw propensity scores are often abstract values that lack clear context, while probabilities are easily interpretable as the likelihood of an event occurring. This makes it easier for stakeholders and end-users to comprehend the model's predictions and build trust in its results.

### Uncertainty quantification

Predicted probabilities convey the model's level of certainty about its predictions. Models can provide probability scores ranging from 0 to 1, indicating their confidence in the predicted outcomes. This uncertainty quantification is valuable in decision-making, as it allows users to consider the risk associated with a prediction and adjust their actions accordingly.

### Threshold selection

Predicted probabilities offer flexibility in choosing decision thresholds. Depending on the specific application and requirements, decision-makers can adjust the probability threshold for classifying positive and negative outcomes. This adaptability allows for a fine-tuning of the model's behavior to suit the desired trade-offs between false positives and false negatives. For example, in the case of a lead conversion, a profitability threshold can be established by taking into account cost of acquisition for each lead, expected profit from actual customers, and probability of conversion.

### Calibration and reliability

When models are well-calibrated, predicted probabilities accurately reflect the real-world likelihood of events. Calibrated probabilities provide more reliable estimates, leading to improved decision-making and better performance in critical applications.

### Ensemble and stacking

Predicted probabilities facilitate model ensembling and stacking techniques. By combining the probabilities from multiple models, ensemble methods can often yield better overall performance compared to individual models or raw propensity score-based approaches.

## Benefits and drawbacks of raw propensity scores

Using raw predictions from models instead of calibrated scores has its own set of pros and cons. Let's explore them.

### Pros of using raw propensity scores:

**Simplicity and Speed**: Raw predictions are the direct output of the model without any additional post-processing, making them easier and faster to obtain. In real-time applications or scenarios where efficiency is critical, using raw propensity scores can be advantageous.

**Suitable for Ranking**: Raw propensity scores can still be useful for ranking predictions. Even though they may not directly represent probabilities, they can help rank instances based on their likelihood of being a positive or negative outcome, which is often sufficient in certain applications.

### Cons of using raw predictions:

**Inconsistent Ranking**: Raw propensity scores might not consistently reflect the true likelihood of events occurring. Two instances with similar raw propensity scores might have significantly different actual probabilities, leading to inaccuracies in ranking and decision-making.

**Suboptimal Decision Thresholds**: Using raw predictions often means relying on fixed decision thresholds (e.g., 0.5 for binary classification), which might not be suitable for all applications. This can lead to suboptimal trade-offs between precision and recall.

**Misinterpretation of Scores**: When raw propensity scores are not calibrated, users might incorrectly interpret them as reliable probabilities, leading to misguided decisions and a lack of trust in the model's predictions.

In summary, while using raw predictions can be straightforward and computationally efficient, it comes with the trade-off of lacking calibrated probabilities and potential inconsistencies in ranking. In many real-world applications, calibrated scores are preferred as they provide interpretable probabilities, better decision thresholds, and improved trust in the model's predictions.

## When should you use propensity score percentiles vs. predicted probabilities?

In the field of machine learning, the choice between using percentiles and probabilities depends on the specific context and the nature of the problem being addressed.

### Using propensity score percentiles

**Outliers are important**: If you are concerned about extreme values or outliers in your data, percentiles can be more robust in representing the distribution. For instance, the 95th percentile captures the value below which 95% of the data lies, making it less sensitive to extreme values.

**Data skewness**: When dealing with heavily skewed datasets, especially those with long tails, percentiles can be more informative than probabilities. Percentiles provide a clearer understanding of the data's spread and central tendencies, even in the presence of skewed distributions.

**Interpretability**: Percentiles can be easier to interpret and explain to non-technical stakeholders. For example, stating that a value is at the 75th percentile means that it is higher than 75% of the other values in the dataset.

### Using predicted probabilities

**Well-calibrated predictions**: Probabilities are essential when you need calibrated predictions. If you want reliable estimates of the likelihood of certain events, probabilities are more suitable. Well-calibrated models can provide accurate and informative probabilities for decision-making.

**Decision thresholds**: In many applications, using probabilities allows you to set appropriate decision thresholds based on your specific needs. For instance, in medical testing, you might want to set a higher threshold to prioritize precision over recall, depending on the consequences of false positives and false negatives.

**Uncertainty quantification**: Probabilities are crucial for quantifying uncertainty around predictions. Confidence intervals and prediction intervals can be constructed using probabilities, giving a range of likely outcomes.

**In summary, use percentiles when you need robust representation of data distribution and are concerned about outliers or skewed data. On the other hand, use probabilities when you require calibrated predictions, decision thresholds, and uncertainty quantification, especially in the context of machine learning models.**

**Ready to try Faraday? [Create a free account](https://faraday.ai/signup).**

Propensity scores and predicted probabilities represent distinct aspects of a machine learning model's predictions. Understand the difference and when you should use them.

Propensity vs probability: Understanding the difference between raw scores and probabilities

Churn looks different for every business, but the ultimate goal remains the same: retaining customers. To succeed in this competitive landscape, it is crucial to have [real-time insights](/predict/category/churn_scores) that can identify customers at risk of churning and engage with them based on their specific needs and preferences.

But where should you start?

Join us in this interactive workshop where we will demonstrate how Faraday's AI capabilities can be leveraged to accurately predict customer churn. By staying ahead in your [churn prevention](/blog/churn-prevention-machine-learning) strategy, you can take proactive measures to retain valuable customers.

**From this workshop, you'll learn**:

- Why modern brands are investing heavily in customer retention
- Ways to reduce churn throughout your customer lifecycle
- Important considerations to effectively predict and prevent churn
- How to get churn predictions into your stack using Faraday
- Activation and engagement ideas to consider

Join Faraday's Directors of Accounts Alex Noyes and Nick Haggerty, as well as Director of Marketing Perry McDermott, where they dive into why brands are investing heavily in retention, and how to do it effectively.

Using churn predictions to improve retention rates

In this webinar, Co-Founder & CEO, Andy Rossmeissl and Director of Marketing, Perry McDermott, walk through some of Faraday's latest features. For a dated list of all of Faraday's product changes, check out the [changelog](https://faraday.ai/changelog).

**From this webinar, you'll learn**:

- **Deployment filtering**: How you can filter & customize [deployments](https://faraday.ai/docs/abstractions/deployments) to target specific propensity scores and personas.
- **Deployment structure**: How to rename, reorder, hide, or structure a deployment to pre-format for specific ad platforms.
- **Managed connections**: How to get started configuring managed connections to your favorite martech software.
- **Score probability**: How a recent update in Faraday used model calibration to improve customer prediction accuracy through [probability scores](https://faraday.ai/blog/faraday-probability-scores).

In this webinar, Co-Founder & CEO, Andy Rossmeissl and Director of Marketing, Perry McDermott, walk through some of Faraday's latest features.

New June features & updates from Faraday

A smart [lead scoring strategy](/blog/predictive-lead-scoring-b2c) is key to ensure you're not wasting time on bad-fit leads that were never going to convert anyway. While it's possible to develop a smart lead scoring strategy manually, it's time-consuming and leaves room for human error. With a predictive lead scoring strategy built by both your data and Faraday's permissioned, opt-in consumer data on nearly every U.S. adult, all the time that would've previously been spent building and maintaining the strategy manually can instead be spent going after best-fit leads.

Join Strategic Account Manager, Nick Haggerty, and Director of Marketing, Perry McDermott, as they walk through how you can utilize Faraday's predictive lead scoring capabilities in your acquisition strategy.

**From this workshop, you'll learn**:

- Benefits of predictive lead scoring for B2C
- How data influences predictive lead scoring strategies
- How to set up lead scoring pipelines in Faraday
- Best practices for planning and implementation
- Activation tips and real-world examples

Join Strategic Account Manager, Nick Haggerty, and Director of Marketing, Perry McDermott, as they walk through how you can utilize Faraday's predictive lead scoring capabilities in your acquisition strategy.

Using predictive lead scores to improve conversion rates and reduce CAC

Faraday’s data science team has made significant improvements to the accuracy and interpretability of customer predictions within the platform; we’re going to give you a peek under the hood at how Faraday predictions are made.

The data science team introduced a process called **model calibration**, which measures how well the model's predictions match the actual outcomes. This method of calibration results in **probability scores**; predictions that are even more accurate and easier to interpret. Let’s take a look at a model report to break down the changes.

## Lift and precision before model calibration

In the below **pre-calibration** graphic, note the large dip in the lift curve (red arrow). With this lift curve, the top deciles of propensity scores, or the people you’d generally be targeting with a marketing campaign based on these results, have lower lift in the population (green square), indicating that there’s room for improvement.

A lift curve with strong predictability should show a more gradual dip than you see below. Similarly, the positive attainment cohort _home and living_ (blue square)–or the outcome you want to predict for–is very close to 50%, which could also be improved through model calibration. This would result in pipeline deployments having, say, a Jane Doe with a 0.8 score that could be interpreted as _pretty likely_ to achieve the outcome.

[![Image of Faraday model report lift table pre-calibration](/images/blog/probability-pre-calibration.png)](/images/blog/probability-pre-calibration.png)

_The above graphic displays a predictive [model report](/blog/faraday-model-reports) pre-calibration. Model reports can be found in an outcome's analysis view_.

Because Faraday uses multiple predictive models to create an outcome, it makes more sense to compare scores across models rather than treat scores from different models equivalently.

## Lift and precision with model calibration

Now, with model calibration implemented in the below graphic, the lift curve’s dip (red arrow) is much less significant for the top deciles, resulting in an outcome that has a higher positive attainment (blue square) in _home and living_.

[![Image of Faraday model report lift table post-calibration](/images/blog/probability-post-calibration.png)](/images/blog/probability-post-calibration.png)

_The above graphic displays a predictive [model report](/blog/faraday-model-reports) post-calibration. Model reports can be found in an outcome's analysis view._

So what does this mean? With probability scores, the groups of your prediction’s population–the deciles–are ranked in decreasing lift values, resulting in higher lift in your top attainers, or the people most likely to achieve the outcome.

With this update, Faraday’s predictions give you probability scores that are directly correlated to the probability of an individual achieving your outcome. If Jane Doe in your population has a 0.8 probability score, there’s an 80% chance that she will achieve the outcome.

Marketing campaigns have never been so predictable.

**[Create a free Faraday account](https://faraday.ai/signup) today.**

How model calibration helps Faraday improve customer prediction accuracy through probability scores.

Faraday probability scores: Improving prediction accuracy through model calibration

Google [just announced you can use Bard in BigQuery](https://cloud.google.com/bigquery/docs/reference/standard-sql/bigqueryml-syntax-generate-text) to do Generative AI at scale. That means you can run your queries right inside of your data warehouse, where your data lives. No more rate limits, HTTP clients, error messages, etc.

Well, it's not exactly [Bard](https://bard.google.com/)... it's actually [PaLM 2](https://ai.google/discover/palm2/), called [`text-bison` in Vertex AI](https://cloud.google.com/vertex-ai/docs/generative-ai/learn/models) land.

It works as "one shot" prompts - where you give it context inside of the prompt itself. It won't remember what you tell it, so it's **safe to use on biz data**. You can use [Vertex AI API to fine-tune your model](https://cloud.google.com/vertex-ai/docs/generative-ai/models/tune-models) and I'm expecting this to be available via SQL soon.

First, create a US-based dataset: (or use one you already have... or another region that supports LLM)

```bash
bq mk --dataset --location us --project_id=$PROJECT_ID $DATASET_ID
```

Next, create a BigQuery connection: (there is no way to do this in SQL as far as I know)

```bash
$ bq mk --connection --connection_type=CLOUD_RESOURCE --location us --project_id=$PROJECT_NAME
Connection 78793318712.us.40833e7c-1c99-440f-911b-c553cf8485f1 successfully created
```

Next, see what service account is associated with that connection: (look at the end of the table)

```bash
$ bq show --connection 78793318712.us.40833e7c-1c99-440f-911b-c553cf8485f1
Connection 78793318712.us.40833e7c-1c99-440f-911b-c553cf8485f1

                          name                           friendlyName   description    Last modified         type        hasCredential                                            properties
 ------------------------------------------------------ -------------- ------------- ----------------- ---------------- --------------- -----------------------------------------------------------------------------------------------
  78793318712.us.40833e7c-1c99-440f-911b-c553cf8485f1                                25 Jun 17:35:01   CLOUD_RESOURCE   False           {"serviceAccountId": "bqcx-78793318712-abcd@gcp-sa-bigquery-condel.iam.gserviceaccount.com"}
```

Next, give the service account associated with that connection access to Vertex AI:

```bash
gcloud projects add-iam-policy-binding $PROJECT_ID \
  --member='serviceAccount:bqcx-78793318712-abcd@gcp-sa-bigquery-condel.iam.gserviceaccount.com' \
  --role='roles/aiplatform.user'
```

Next, create a remote model inside of BigQuery: (use the connection ID you got above, and note that I'm naming the model `my_llm_1`)

```sql
CREATE OR REPLACE MODEL
  `$PROJECT_ID.$DATASET_ID.my_llm_1`
REMOTE
WITH CONNECTION `projects/$PROJECT_ID/locations/us/connections/40833e7c-1c99-440f-911b-c553cf8485f1`
OPTIONS (remote_service_type ='cloud_ai_large_language_model_v1')
```

Then you should be able to run LLM queries:

```sql
SELECT
  *
FROM
  ML.GENERATE_TEXT(
    MODEL `$PROJECT_ID.$DATASET_ID.my_llm_1`,
    (
      SELECT
        "Say hello world" AS prompt
    ),
    STRUCT()
  )
```

But the killer app for LLM in the data warehouse is to **run prompts across tables full of data**. Here's how I tried sentiment analysis for some financial products:

```sql
CREATE TABLE `$PROJECT_ID.$DATASET_ID.my_outputs` AS (
  SELECT
    my_id,
    ml_generate_text_result["predictions"][0]["content"] AS sentiment_analysis,
  FROM
    ML.GENERATE_TEXT(
      MODEL `$PROJECT_ID.$DATASET_ID.my_llm_1`,
      (
        SELECT
          -- return my id so I can join this back to my data
          my_id,
          CONCAT(
            "Please provide an entity sentiment analysis of bonds, annuities, and CDs in the following text: ",
            text_content
          ) AS prompt,
        -- here's the magic: running a prompt for every row of `my_inputs` table
        FROM `$PROJECT_ID.$DATASET_ID.my_inputs`
      ),
      -- tell the LLM not to be super creative, but just return the highest probability results always
      STRUCT(
        0 AS temperature,
        1 AS top_k
      )
    )
)
```

Here's the final production query that uses some regex parsing to extract structured data from the LLM response:

````sql
CREATE TABLE `$PROJECT_ID.$DATASET_ID.my_outputs` AS (
  SELECT
    my_id,
    CASE
      -- if it gives back raw JSON, use it
      WHEN LEFT(JSON_VALUE(ml_generate_text_result["predictions"][0]["content"]), 1) = '{'
      THEN SAFE.PARSE_JSON(JSON_VALUE(ml_generate_text_result["predictions"][0]["content"]))
      -- if it gives back JSON inside a markdown block, use it
      ELSE SAFE.PARSE_JSON(REGEXP_EXTRACT(JSON_VALUE(ml_generate_text_result["predictions"][0]["content"]), r'```json([^`]+)```'))
    END AS sentiment_analysis,
  FROM
    ML.GENERATE_TEXT(
      MODEL `$PROJECT_ID.$DATASET_ID.my_llm_1`,
      (
        SELECT
          my_id,
          CONCAT(
            "Please provide, as a JSON object with only entity and sentiment, an entity sentiment analysis of bonds, annuities, and CDs in the following text: ",
            -- remove backticks that might interfere with parsing response
            REGEXP_REPLACE(text_content, r'`', "")
          ) AS prompt,
        FROM `$PROJECT_ID.$DATASET_ID.my_inputs`
      ),
      STRUCT(
        0 AS temperature,
        1 AS top_k
      )
    )
)
````

I'm hoping it will get easier to extract structured results soon, using [jsonformer](https://github.com/1rgs/jsonformer) for example.

Thanks to the [BigQuery ML team](https://cloud.google.com/bigquery/docs/bqml-introduction) for implementing this. I think it's the obvious place for LLM!

Google just released LLM support inside of BigQuery. Here's how to start using it today.

How to use Google Bard in BigQuery (quickstart)

## What is market sizing?

Market sizing is a method of calculating a company’s revenue potential in a geographic market, which can provide crucial insights for competitive brands to be aware of when looking to expand, otherwise resources are likely to be wasted.

Market sizing is not a new concept, but more recent innovations in data collection and predictive analytics solutions have opened the door for more accurate market size calculations, enabling brands to make better go-to-market decisions and expand with more confidence and direction than before.

## Top down vs bottom up market sizing analysis

There are two methods of market sizing analysis: **top down**, and **bottom up**. In this article, we’ll look at both, and then explore how using predictive analytics for bottom up market sizing can provide you more accurate estimates while drastically expediting the analysis process, and give you actionable insight into where you’re most likely to have success penetrating.

### Top down market sizing

When using top down market sizing to calculate your market potential, you start at the highest level of the market that you’re in, where you find the largest possible size of the market you could acquire, then multiply it by the percent of the market that your product fits in.

For example, say you’re a brand that sells humidifiers, and you’re planning to launch a new humidifier that’s perfect for maintaining moisture levels in a newborn’s nursery. With the top down approach, your research shows you that the US market for humidifiers is \$700 million, and 65% of humidifiers sold annually are single-room, like your new product.

> **\$700m X 65% = \$455m**

Since, for the most part, the numbers that are relevant to your company and industry are likely already published elsewhere, this approach involves doing some research to find the populations and market values you need rather than calculating them yourself.

**Pros**:

- **It’s fast**. The research involved and the calculation are both straightforward and quick, so there’s not much time investment in a top-level market size estimate.
- **Accessible data**. Regardless of the outlet you want to source your data from, people have been publishing industry reports and analyses for years, so the data you need can easily be found via some time on Google.
- **Helpful at-a-glance**. Since top down market sizing looks at broad trends and market-wide data, it’s useful to be able to know, at a glance, if the space you’re looking to go-to-market in is worth the investment you’d have to make.

**Cons**:

- **It’s very broad**. As mentioned above, it’s useful to run a top down market size analysis for an initial glance at the market, but it’s unlikely to be specific enough for your company’s needs.

### Bottom up market sizing

Bottom up market sizing starts on the opposite end in that it begins with your niche. Using the same humidifier example, you’ll likely want to find people with children in the household to target since your new humidifier is for nurseries.

Through extensive research, data collection, and surveys, you find that there are 4 million people in this target group. To calculate the bottom up market size, you’d then multiply the number of people by the annual revenue you’d expect from them for the product–your customers are unlikely to buy more than one humidifier in a year, and its cost is \$100 per unit, so we’ll use \$100.

> **4 million people X \$100 = \$400m**

So what’s the verdict? Top down or bottom up market sizing? Since it uses data that’s relevant to your company, bottom up market sizing provides a more accurate estimate of your potential market–if your company has the bandwidth to put in the time. Fortunately, predictive analytics can help cut down that long timeline dramatically while providing you with geographies where your most likely buyers live.

**Pros**:

- **Backed by your data**. Since bottom up market sizing uses your own data & analytics through your customers, product, and cost data, results are much more tailored to your company, and they’re more realistic to boot.
- **As niche as you want it**. Rather than relying on industry-wide assumptions, this strategy begins and ends with your company’s niche.

**Cons**:

- **Time-intensive**. Performing the data collection required to accurately run a bottom up market sizing analysis is a long process in that, whether you’re conducting interviews or surveys, you’ll need to wait for results to come in.
- **Resource-intensive**. Assigning teams to create surveys, conduct interviews, act on their results, and so on can occupy a significant portion of their time.

## Calculating market size with predictive analytics

In the end, bottom up market sizing wins out as a more effective approach to calculate market potential. The downsides, of course, are the longer timeline and additional resources required to do it effectively.

This is where predictive analytics comes in–leveraging machine learning to predict individuals’ propensity to buy your products or services. With machine learning, your time-to-value in performing a bottom up market sizing analysis is dramatically reduced because it eliminates the need to design surveys and conduct interviews, conduct research, and wait for results.

Better yet, rather than spending extra time using your analysis to plan where you _think_ your campaigns will have the most impact, likely buyer scores show you _exactly_ where your high-intent buyers are at any geographic zoom that suits your needs.

### Traditional approach to bottom up market sizing

As we reviewed earlier, the typical traditional approach involves multiplying the number of people in the segment by the annual cost of your humidifier, which is simple enough math once you’ve done the required research to find the number of people in the segment.

At the end of the day, there’s still a lot of guesswork involved. In your research, you uncovered the raw number of potential buyers in your market, but the only direction you might have as for where to direct your campaign are surveys, which are unlikely to have reliable penetration for guidance, and assumptions–maybe the 5 driest states in the U.S. are the best bets to market your humidifier?

### Predictive approach to bottom up market sizing

With the predictive approach, you’ll take the segment you created above, then use AI to predict the propensity to buy for everyone in the segment based on real historical data. Machine learning algorithms surface patterns in the data to calculate just how likely each individual in this segment is to buy the product in likely buyer scores, and their forecasted spend over their lifetime as a customer of yours.

These likely buyer scores can be filtered by county, state, city, or metro, so that you have crystal clear indicators where your high-intent buyers are. You can even use your own points-of-interest, like brick and mortar locations, to perform radius-based likely buyer targeting for pin-point accuracy.

With Faraday, the predictive approach can be done in just a few clicks, saving you a significant amount of time and resources, providing more accurate results, and helping align your strategy through clear indicators for the locations you’ll have the highest likelihood of penetration in.

## How to do bottom up market sizing with Faraday

Using Faraday, the predictive approach for bottom up market sizing can be embedded wherever in your stack you need it in less than a day. No developers, no data scientists, and no long research timeline required.

### Connect data and define customers

To start, you plug in your data by mapping your data fields to Faraday properties.

![Image of Faraday dataset](/images/blog/market-sizing-dataset.png)

Then, in cohorts, you’ll utilize both your customer data and Faraday’s built-in consumer data to build the [exact cohort](https://faraday.ai/docs/cohorts/customers_from_hosted_csv)–or interesting group of people–you want. To continue with the nursery humidifier example, you’ll define a cohort for your customers, and a cohort using the trait for _children in household_, sourced from Faraday’s [rich, cookieless, opt-in consumer data](https://faraday.ai/features/data), to target with your propensity to buy predictions.

![Image of Faraday cohort](/images/blog/market-sizing-cohort.png)

### Predict propensity to buy to hone in on likely buyers

With your cohorts defined, build your predictive outcome by choosing your attainment, or your customers. Here, you’re defining your propensity to buy, or the likelihood that anyone in the U.S. is going to become a customer by making a purchase.

![Image of Faraday outcome](/images/blog/market-sizing-outcome.png)

Now that your predictions are defined, create a pipeline to make them actionable by choosing your **likely buyers outcome** and your **children in household** cohort.

![Image of Faraday pipeline](/images/blog/market-sizing-pipeline.png)

Machine learning algorithms then get to work predicting how likely people across the country with children in the household are to buy your new nursery humidifier. You can even further your bottom up market sizing analysis by generating customer personas and their predicted lifetime value with your brand.

### Deploy likely buyer scores to your analysis tools

Finally, deploy your likely buyer predictions back to your stack, via CSV file for versatility, or right back into your database or data warehouse for analysis in your BI tools. When deploying, filter to whatever aggregation level makes the most sense for your strategy: county, state, city, or metro. In this example, perhaps choosing state will reveal that the driest states in the country have the highest propensity to buy individuals.

![Image of Faraday pipeline](/images/blog/market-sizing-deployment.png)

With the predictive approach to bottom up market sizing, you have access to likely buyer scores anywhere in your stack in just a few clicks, providing insight into what geographic locations your company has the most high-intent buyers in so that your market size is more than just a rough estimate–it’s actionable and data-backed.

**Ready to try it out? [Create a free account](https://faraday.ai/signup).**

What is bottom up market sizing? What are its benefits? How can B2C brands accurately analyze markets without investing vast resources?

Bottom up market sizing: how to calculate market size using predictive analytics

Founded in 2016, [Ethos](https://www.ethoslife.com/) has been on a mission to make life insurance easily accessible to everyone by providing a faster and easier way to buy life insurance online. How do they do it? By automating the underwriting process to identify the right products for users in real time. This means less time spent going back and forth with different agents and carriers to find the products that will work best for them.

However, despite their success, Ethos knew that with an improved understanding of their customers, they could do even better. We met with the data science team at Ethos to learn more about their marketing strategy, data challenges, and how the team is using Faraday to address them.

## Informing marketing and product strategy using Faraday personas

By tapping into Faraday's [built-in consumer data](https://faraday.ai/features/data) and machine learning engine, Ethos identified four [distinct personas](https://faraday.ai/predict/category/persona_assignments) in their customer base and assigned them to customers in their [database](/integrations). After analyzing different metrics like conversion and churn rates across these core segments, Ethos identified two promising groups which helped inform important adjustments to the company’s marketing and product strategies.

## Reaching 50% higher ROAS with innovative targeting strategies

With recent changes to Facebook’s advertising functionality, Ethos needed to find a way to navigate around a crowded platform while trying to optimize their return on ad spend (ROAS). So the team experimented with an innovative targeting strategy: using customer predictions from Faraday to seed a new Facebook Lookalike Audience.

The persona-seeded audience resulted in significant ROAS improvements—50% higher than equivalent campaigns that used broad targeting strategies—so they took it a step further by adopting these audiences across the board.

## What’s next for Ethos?

Ethos is now conducting more qualitative research around the core customer segments they identified with their Faraday personas, which they’ll use to further personalize experiences for their customers. Additionally, they’re planning to use customer predictions from Faraday to optimize their targeting strategies across other marketing channels.

We love working with [Ethos](https://www.ethoslife.com/)–they’re on a mission to make life insurance accessible for all, and with Faraday, they’re finding innovative new ways to accomplish that mission.

[Create a free Faraday account](https://faraday.ai/signup) today.

How Ethos uses Faraday to speak their customers' language and optimize targeting strategies.

How Ethos uses AI to streamline the insurance buying experience

_This customer story has been anonymized to respect the company’s competitive strategy._

This customer story explores how a leading home improvement brand used Faraday’s customer prediction platform to identify and engage more likely buyers and score their extensive lead database, optimizing their ad targeting strategies and lead conversion rates simultaneously. Here’s how they did it.

## Ad targeting strategy

The home improvement brand wanted to [reach likely buyers on Facebook](https://faraday.ai/predict/pipelines/likely_buyers_in_facebook). Using Faraday, they built predictive models based on a cohort of customers acquired through paid social channels and leads generated from their online scheduling platform, enabling them to identify likely buyers and individuals actively interested in home improvement services across the US.

### Scoring and audience selection

Faraday enabled the brand to score the entire US population using the custom models. Each individual received a score reflecting their likelihood of being a high-potential buyer or lead based on the models' predictions. Top-scoring individuals were then selected to create two distinct audiences: one from the customer lookalike cohort and the other from the leads generated through the online scheduling platform.

### Facebook ad campaign success

With their target audiences defined, the brand launched a comprehensive Facebook ad campaign and tailored the creative content and messaging to resonate with these audiences. **The campaign yielded outstanding results: 35% increase in incremental lead generation and 14% increase in their average close rate.**

## Predictive lead scoring

The brand's CRM team is responsible for managing and nurturing leads acquired from multiple sources. With a diverse lead pool, they needed a more robust and scalable solution to effectively prioritize leads, so they leveraged Faraday for predictive lead scoring. The team integrated their lead data with Faraday and built [predictive lead scoring](https://faraday.ai/blog/predictive-lead-scoring-b2c) models to assess and rank leads based on their likelihood to convert into paying customers.

### Third-party validation and results

To validate the accuracy and effectiveness of [Faraday's lead scores](https://faraday.ai/predict/category/lead_scores), the home improvement brand engaged a third-party agency to conduct a thorough analysis of the predictive lead scores, comparing Faraday model predictions against actual lead conversions and outcomes. **The analysis revealed that the conversion rate of leads in the top 50% of lead scores was 4x higher than the brand’s baseline conversion rate, confirming the model’s ability to accurately identify high-potential leads.**

## What's next?

Faraday's customer prediction platform enabled the home improvement brand to identify and engage more likely buyers and high-potential leads, improving lead generation and conversion rates simultaneously. Building on this success, the brand is now exploring their customers’ preferred communication channels and plans to implement [repeat purchase predictions](https://faraday.ai/predict/category/repeat_purchasers) and lead revival scoring to re-engage leads that cancel appointments.

[Create a free Faraday account](https://faraday.ai/signup) today.

Faraday enables this home impovement brand to power lead scoring and acquisition efforts resulting in real, third-party validated results.

All posts