Using machine learning to predict the leads that close | Hacker News Reader

Using machine learning to predict the leads that close

outfunnel.com

57 pointsstandrews30 comments

What should you do with this information though? Should the salespeople focus on the customers that are most likely to convert? Or should the salespeople give minimal attention to those, and focus most on the ones less likely, but not 0 probability, to convert.

I think the most important outputs of this are understanding the factors involved in conversion to tune business processes, not necessarily using the outputs of the model to target specific users.

Yeah, the salespeople avoid spending time on leads unlikely to close.

They're all operating on this kind of information already (like the basic stuff - is this person a decision-maker, do they have budget etc etc).

I take this seriously because I've seen consultants making money in this space already (advising on leads unlikely to close).

Maybe it’s like the nuance in music - knowing the notes not to play. I think the art is in weighing the variables as a human because they can change also based on competition in the room. Having a career ongoing in sales support I’ve seen first hand how erratic the decision trees can be for private or public organizations. While the general setup is similar “do business with X with Y” the ingredients can differ widely to get to the sale. ML might tell you to send a holiday gift but I bet the human has a better idea of what kind of gift to send than ML.

I think I agree, and we could make exactly that observation about syntax highlighting. It isn't like syntax highlighting is telling me things I already know, but it is a free win for my productivity while coding.

Making something that is already obvious cheaper to discover can help the people who need to act on it.

But then this has the impact of reenforcing model predictions (unless controlled for). This lead won't convert based on the model, so I won't spend time with them or give them any preferential pricing, so the lead doesn't close (and the converse is true).

This isn't a pure ML problem, and without "treatment" data I'm not quite sure how the blog is adjusting for customer propensity towards an outcome :/

Look up the guidelines that emergency responders use when tending to incidents with a large number of casualties. The techniques are just as relevant to leads conversation.

Also, explore the BANT leads conversation ideas. Adapt it to your industry.

This is not like medical triage where some people will just recover on their own... there are vanishingly few customers who are going to buy without a focused sales effort. Your time is best spent on those most likely to convert because even "most likely to convert" in B2B sales is not someone coming to your office and pounding your desk and saying TELL ME WHAT I NEED TO DO FOR YOU TO LET ME BUY THIS

Many B2B companies kidna bring this on themselves by hiding pricing and product information behind a sales wall, forcing you to do a little dance with them to get basic information without seeming too committed.

You could charge more or less and optimize revenue

The answer here, as always, is "it depends".

It almost always makes sense to not waste time on "dead" leads. Best to start there.

You'll probably get the highest ROI when you focus on the leads that are somewhat likely to convert (ie. you'll influence those sitting on the fence).

But if you're short of people and have lots of high-quality leads, you'll probably want to focus on the leads most likely to convert.

Back at my previous company (fast-growing SaaS), we ran an AB test where leads were split into six groups: half received sales touches and half didn't, and there were three lead score groups in both (low, medium, high probability to close). Working with the medium group gave the biggest uplift.

“Working with the medium group gave the biggest uplift.”

Anecdotally, I’ve seen the same thing in a B2C context. The uplift in the highest probability group was so bad that we would leave those leads alone completely, even though the marginal cost of an email or sms is basically 0 as a % of revenue from a successful conversion.

I worked previously at a company that was trying to do a predictive analytic, ML on top of Salesforce (to qualify and score leads, generate new promising leads, etc). The task was extremely difficult and from what I saw people were getting some marginal improvements from ML itself. Most companies at the end were interested in basic (but quite hard things to do right) - cleaning, dedupoing, enriching, generating more leads based on ICP, etc.

Not sure if in this case it's something different but my take from the past experience it's hard since data is noisy + closing the deals depends on a human as well and these tools don't take that into account usually.

I hope though there will be an ML product that can do a decent improvement in this space. Even small improvement there brings significant ROI.

I spent years using a tool called Clari that applied analytics to Salesforce and Office (email) to predict deal activity. When it worked, it was really good.

But it had a fatal flaw: relying on Salesforce meant relying on the data that sales people input. And I quickly learned that sales people hate reporting tools, update them only under duress, and generally fill the database with crap.

Perhaps it's different when you are selling subscriptions over a digital channel, but for classic B2B feet on the street deals, ugh.

Thus spawning a whole new class of tools that add a more convenient interface on-top of salesforce.

Right, but garbage in, garbage out, whether you are feeding your CRM directly or via a shiny UI layer. No amount of ML can save you if the sales droid isnt filing data.

Can’t track what you don’t know about is a constant issue with sales groups lacking proper structure and processes - it should be easier to do it right than to sidestep the resource.

> sales people hate reporting tools, update them only under duress, and generally fill the database with crap.

Sometimes the salespeople do not want to share information about their leads, as their over many years tediously-spun network of their private connections and business-friends is their most valuable asset.

Sales is two-ways -- need to get get sales credit and keep reputation.

Looks like that company needs to remind the sales drones that coffee is for closers, second prize is a set of steak knives, and third prize is you're fired.

Every people hate reporting tools.

Spending time and energy on describing the job you've done yesterday is taken from the job you are doing now. Not only it is unfulfilling work, but it's also often not taken in consideration by highers up that will likely assume your regular job should remain unaffected by reporting.

Then you get the feeling on being spied upon, nevery good for trust or moral.

And finally, you know that some actions you take will be judged by the people reading the reports. Unfortunately said judgment as ha chance to be unfavorable yet unfair, because the person making it might not have the context, personality, or knowledge required for a fair one.

However, workers have little chance to gain anything from favorable reports. So the asymmetry is just very much against the person filling the reports, who is paying the price for it.

What do you expect?

I like your blog on how you use EDA, but I'm not sure I'm getting the Machine Learning piece. It would be nice if you guys went into more details, but I appreciate how you were able to tie together different data sources and walked ppl through the analysis!

I am guessing they used logistic regression because they could show the lead win probability as a function of emails opened. It's quite easy to do with logistic regression.

nah its ml bro

There's a separate technical post coming - will post a link here once it's live.

In Shakespeare's Macbeth, the witches prophecise to him that he will become King. Spoiler alert, he does, by killing the previous King in order to make the prophecy come true.

Which sales leads to focus on? We were in a unique position to answer this because we connect data from marketing tools like MailChimp with CRMs like Pipedrive, plus we track web visits.

Why does the X-axis on the marketing email count graph to up to 60? Is this where we've ended up, sending 60 emails to the same lead? Ouch.

It's someone opening or clicking emails 60 times (can happen!), not sending 60 marketing emails.

With Gmail and the like you only get to know about the first open.

I wonder if this is actually caused by the email being forwarded around to many people.

Email forwarding or just quirks of various email services and clients. Email clicks or website visits definitely a more reliable signal

Reply on news.ycombinator.com