20th: Important -Financial- Decision Maker — Part Two

In part one, I came out with the idea of building a simple training model for making important financial decision. Later I rejected my own idea, because we can’t make financial related decision purely using data, and disregard emotional factors, which actually play much higher weight in the decision making process.

After all, human should not rely on machine to tell them what to do. Human should learn from their own mistakes by doing it.

I learn that from Simpsons 

This episode is about Lisa and a group of female programmers building an application that tells the user if it is appropriate to post the thing to social network the moment the user tries to post. In the end, the robot convinces Lisa he is real and he is under too much stress to tell people if they should post or not, because most people are trying to post stupid things, which has the obvious answer without even the need of consulting a robot.

I wasted too much time on watching Simpsons, but surprisingly learned a few interesting life lessons. This is one of them.

What about Crowd Sourcing Decision Making?

Crowd sourcing to make decision is not really new.

Recommendation can be seen as one of the designs. However, it is easy to recommend similar items to buy, or restaurant to go, or place to visit, because they are largely supported by data with high probability. More importantly, recommendation has no responsibility to take.

Financial decision making is different. The difference is basically:

do I buy a car, v.s. which car should I buy?

A decision like this involves emotion and rationale, which cannot be calculated (accurately) by machine.

In fact, someone has already asked if such a thing exists.

Surely, there are some products out there:

All very different. If you look at the transparentchoice, which has an example of buying a car. It looks like it’s trying to buy car 1 or car 2. Anyway, it’s fairly complicated and not easy to understand.

How will I design it?

First, ignore how practical the idea is. How will I decide it?

Since it is a crowdsourcing site, it should have the design that encourages strangers to engage. This is generally done by materially (incentives) or morally (social class). Giving out incentive is probably not possible, neither is the other option, as we are not starting from Facebook or Twitter. One technique people use is sympathetic, which is putting visitors in the same shoes of the person who asks the question, thereby motivating the visitors to complete the request, since he or she feels it’s like doing his/her own favor.

With a quick sketch, I came out with the above design.

  • the visitor immediately enters a decision making procedure (omit the landing page)
  • the visitor is presented as a single question: being helpful
  • the visitor can see how similar himself or herself is, compared to the requester
  • the visitor can see helpfulness points progression

One part of the page (the green button) will continue asking questions based on the previous answer. Those questions are provided by the requester, or suggested by some machine learning.

In this flow, the visitor never needs to answer YES or NO to the real question, because that will defeat the purpose of learning and crowd sourcing the decision making. Therefore, on the second thought, maybe the web page should hide the real question, which is “Should I buy a car?”.

Maybe the landing page that leads to this procedure will be based on proximity or ages, which makes the visitors feel closer to the requester.

In the end, the visitor will gain the helpfulness points (as an incentive), which varies by the number of questions he or she has answered.

Will it work?

Slowly, this website will accumulate people who have answered many questions and gathered the answer data for groups, like people who live in similar areas, same genders, same age group, work profession, etc.

This will help us to build  a data set helps the decision making.

But, there is still no emotion or rationale factor.

Is it though?

This is actually a fantastic human study topic. God, I miss grad school when I can just came out with a hypothesis and start doing a user experiment.

The difference is how we gather the data.

In part one, I was initially thinking about hard data, which are numbers related to financial situation: income, assets, debts, etc. Those will work, but they are too naive for decision making.

In the approach I just outlined, it’s based on similarity of the requester and the helper. And the question is never about hard number (or maybe we should still weight those in, depends). For example, questions to ask in the decision “Should I buy a car” can be:

  • Do you have a car already?
  • Is your car working?
  • Are you comfortable with additional cost each month, or just pay a big lump sum now?
  • Is it stupid to buy a second car but don’t drive too much?

You see, those are soft data.

If I need a crowd sourcing decision making, I will not care the opinion from a retiree who lives in Texas and has a farm that raise horses as hobby. That is very specific. I will be more interested in the opinions from people who are 25 years old (ssssssssshhhh), live in the same area as me, as well as working in the same profession.

Therefore, I think it will probably work.

I dibs this idea: a crowd sourcing website for decision making that is based on similarity of the crowds.

19th: Important Financial Decision Maker — Part One


A friend put his cat on life-support for over $10,000. And his cat survived after nearly 2 months.

Amazing story, but an debatable decision, as he still carries student loan, pays rent, and still tries to save more.

No one in any position to judge this decision is right or wrong, since everyone has different feeling to different things. Something may be worthy to someone, but may be given up by a different person.

Machine Learning for making Financial Decision

If you think about financial decision, it is fundamentally a classification problem, which is a much easier problem to solve compared to deep learning or Neural Networks. A simple classification problem can be seen as:

We provide a set of data and we say they always yield the result “Class A”, and then provide a different set of data and we say they always yield the result “Class B”. Next pick a function (linear or non-linear) to train with those 2 set of data. Under the hood, it’s going to refine some parameters of the function. Think about “ax + by = C”. The idea is to get the parameter “a” and “b” as best as they can.

After that is done, now feed the function with a new set of data, which will yield a result of either class A or B. With additional cross-validation, some training model can have a probability to support the result. Sometimes a classification problem may yield multiple classes. Generally, this kind of solution yield better results with more data.

With this superficial introduction of one kind of machine learning, now if we look at the financial decision problem, isn’t it just a simple 2-class problem:

  • Do it, or
  • Don’t do it

Training Data Set

As long as we can remove emotion from the equation, this should not be too complicated.

To prepare our training data, we can simply list some hard data, including:

  • debts
  • assets
  • income
  • life-time goal

Someone needs to calculate and determine a weighted point with each pair of those data. For example, if debt is higher than income, a higher weight for NO-class is given. And then we can (probably) easily train a model for making financial decision.

Now, if I want to know if I should buy a new car. By providing my current financial situation, this application should tell me a YES or NO question. Of course, something like buying a car may require additional data, like do I have a functional car already, and how much do I like my car?

The latter is tricky.

Why there is NO such a thing?

If data can really help us solve this problem, why there is no such a thing?

First, emotion or feeling is tricky and impossible to calculate.

At least in most case.

If someone has no income but rather pay a high pet medical bill with credit card, that is not really applaudable. But if someone has stable income and feeling to the thing he or she is paying for, then there is no formula can be used to calculate.

Second, human has rationale. At least for most people when they are not drunk.

I have a car now. It is very old, 2005 Element. In summer time it has a squeaky sound that drives me nuts. I have been thinking to get a new car, but my current car is still functional and I also have spent a lot of money to fix its old problems. Financially I can support myself to get a new car, however, I really see no strong reason to get a new car at all.

How can we rely on computer to factor emotion and rationale? At least I have no knowledge how this can be achieved.


This time I kind of have an idea, but honestly this is not a useful idea. Rather, it’s like a thought.

Can neural network or deep learning really calculate human feelings and rationale? I do not know. But at least, I don’t think data can solve it.

To make important financial decision, we probably still should rely on human sense. Of course there is finance adviser, but they just want to make money from you. So always rely on yourself for important financial decision and trust no one.

18th: Moodger – A device may change your mood

I started with “A device that changes your mood”. And then I paused a while. It sounds like a medical device and may even feel a little deceiving, so I later change it to “A device may change your mood”.

This is not a lie. It is just another useless idea.

Is it though? Hear me out.


Yes, most people know about Nest.

It’s a device that connects to the thermostat at your home, and lets you remotely control the temperature. Big deal, right? What makes people fall in love is that it will learn from human input and eventually starts adjusting itself.

In addition to that, it also has some nice features for energy efficiency. It generates a report, of course, every one loves reading a report and feel good about how much energy they have saved after using the device.

I have no objection that features in Nest should be considered as parts of the future smart home. Nowadays, we have all sorts of “smart” devices at home: smart light bulbs, smart monitors, smart fridge, smart microwave, the list goes on.

Yes, people in the future might really want to tweet from their fridge. But I am not falling in the same trap, only staying in the fixed mindset of “connection with machines”. Human, we should connect with human, device that connects people, that is the direction I want to go.

What brings human together?

  • money? hell no
  • food? maybe
  • the answer is laughter!
    • a smile on anyone’s face brings a stranger closer
    • an innocent laughing can even melt the vicious enemy’s heart

If you don’t believe me, try this:

A brief history

Many years ago, I wrote a small program called Jok4U. It can still be found in Wayback Machine. The idea was it randomly popped up a joke in your computer. It was a very naive program and did not do too much. But I tried…

A modern version

Now what about this?

Instead of having a program installed in your computer, what if there is a small device, just like Nest, which is mounted on a wall. When you pass by, it will randomly ask you if you want to hear a joke. Something like what the following:

That user figure is passing by the device. The device has a motion sensor, so it knows someone is nearby. It asks “hey, do you  to hear a joke?”. If the user answers, “yes”. The device will tell a joke.


Feature List

There are more, of course! The device support:

  • can be configured to have different level of jokes: kids-friendly, safe for work, or adult-only
  • can learn from the human answer, eventually it will automatically tell joke instead of asking question first
  • can learn from the human response and adjust the joke selection algorithm
  • can learn from the human response and adjust the way of telling jokes, it will start with the cold machine-like voice, but slowly, it will adjust itself using Adam Sandler or George Carlin’s voice.

Moreover, we have to have this feature

  • can collect your laughter and then generate a report, which analyzes how much laughter you have and how much calories you have burnt from laughing.

This device will of course have to connect with Internet, so it can also download the jokes accordingly and upload the data for machine learning purpose.

Is it really a stupid idea?

If anyone is reading this idea, let me tell you why this is not a stupid idea, and also let me confirm you that I have been serious the whole time.

Use Case 1: Ease the tension

First, I love my kid.

But I admit some times it may get heated at home, especially, when

  • teaching her math…
  • asking her eat vegetable, 20th times during one meal time…
  • telling her the same thing I have told her 100th times…

many other cases. I told myself I would not yell at my kid before she was born. But what I learn is, yelling at your kids does not make you a bad parent and does not mean you don’t love them (Murray Goldberg may not be a good dad model, but he loves his children).

Sometimes you just do what you have to do.

So, imagine when you are yelling at your kids, the device can sense the pitch hike and say “Knock Knock”. Both you and the kid will probably stop, and answer… “Who’s there?”. Now the tension is cool down.

Use Case 2: Change the mood

Being a grown-up sucks. Work is hard, traffic is bad, greedy banker and incompetent politicians never stop screwing economy up. It is difficult enough for people to laugh more.

What if when you open the door of your home, the device will take an image of your face and analyze your mood, and then respond accordingly.

Using thermal image to understand human emotion is definitely not new, see these paper:

The device will be capable of detecting the user current emotion: angry, or depressed, or sad. The device then reacts different by using different type of jokes, or maybe just play baby laughing. The idea is trying to change the mood of the user at the right time.

I am not a psychology expert, so I cannot say this will really work. More research may be required to validate this intention and use.

Use Case 3: Ice Breaker

Now imagine if you have a party with some guests in the house. The device can detects more than usual amount of people passing. If the human voice level suddenly drops, this device can jump in to break the ice by telling a joke.

I cannot imagine how useful this device can be when a socially awkward person brings a date home. It is like a little wing man.

Who am I kidding, a socially awkward person never gets a date.

What is the name “Moodger” from?

It stands for “Mood Changer”. I actually like this name.

Basically, the idea is building a robot that tells jokes, reacts to human, understands human, and more importantly, alter the mood such that a human can still have a feel with other human as well as the world.

Many people may underestimate the power of laughter.

Building a robot telling joke is not really novel. In fact, back in 2010, researchers had already started this. See this article: Joke-Telling Robots Are the Final Frontier of Artificial Intelligence.

However, the novelty in “Moodger” is not really just a robot telling jokes. It’s a sensible helper. Just like Nest, it will learn from human command, human response, and ambient atmosphere to react accordingly.



13th: Hunter’s Feeder for cats and dogs

I have a Siberian cat. She is beautiful, but like most other cats, she is a pain sometimes. For example, she refuses to eat wet food, unless I hold the wet food in my hands. I also have to try many types of dry food to finally settle down one brand.

I use this auto feeder. It dispenses specific amount of food at a specific time. The timer is defective, always moves slower. Other than that, it works well. In terms of dispensing,  it basically drops the food in a bowl.

PetSafe Healthy Pet Simply Feed Automatic Pet Feeder

Unfortunately, my cat is not really into this feeding style. As I mentioned, she loves to eat when the food is in my hand. But I am not really a big fan of that. She will still eat from the bowl, but usually she will meow a while, see I am not going to feed her with my hand, and then finally she will go eat them. She seems to be more happy when she can eat in her favorite style.

There is, however, another way she loves to eat her food. It’s easier to show what it is than describing it:

Food hunting

Basically, I glide a dry food on the floor, and she would chase it and stop the food, then eat it. It may be difficult to see the food in the GIF.

My cat loves it.

I think deep inside, she still has the hunting instinct. Siberian cats are said to be dog-like and love chasing small animals.

I hate to let her go outside, because of her long fur, which attracts a lot of mud and leaves, very difficult to clean. I guess that is the main reason of me feeling guilty and letting her to eat from my hands so many times.

But you see, I have to stop doing that. Then I have to do something with the other option. Chasing food down reminds me of this:

iFetch Interactive Ball Launchers for Dogs
iFetch Interactive Ball Launchers for Dogs

This is a little machine that launches a ball for dogs to chase.

Yes, we need a similar machine to let cats or dogs to chase food.

This machine, I name it:

Hunter’s Feeder

I am going to learn how to build something like this, but here is a list of features I am thinking:

  • shooting food in at least 2 different ranges, i.e. short distance and long distance.
  • alarm mode: rings to prepare dispensing food so the pet can learn the time is ready.
  • user setting: user can select when and the amount of food to launch.
  • rotating shooting head: the launch angle can be 360 degree.
  • visual detect: the machine will detect if there is any obstacle ahead before launching and automatically select a direction to launch.
  • self-learning: it can learn how to train a pet to start chasing the food further or to a different direction, always keeps the pet interested. Say if a smart cat realizes the machine always just launch food to one direction, she just stays there and waits, instead of coming closer to the machine and then chasing. The machine needs to have algorithm to avoid this behavior.
I don’t have to do it again

OMG, I love this idea.

Now I really need to learn some mechanical engineering.

12th: Another way to buy used cars and more

Here is this week’s idea: why are all websites sells used cars the same?

They all have something like this:

from https://www.autotrader.com

They all have this kind of control for the users to filter what they want. It is a reasonable assumption. This kind of control is used by the potential buyers. However, has anyone thought of: this kind of website is actually initialized for the sellers.

By that, I mean: a seller will have to first go to that website, enter the selling vehicle’s information, as detailed as possible. Of course, it is also possible that a web crawler can do this job. It pulls the information from all sorts of website and extract those data from the description.

Therefore, a website that sells used cars must first have used cars to sell, and then when potential buyers come, they will be able to search.

E-commerce Websites are the Same

In fact, not only used car websites, all other E-commerce websites are the same. They first have things to sell, and then they attract buyers to find them. They all start from the seller’s perspective.

Another Perspective

What if a website starts from buyers’ perspective?

Let’s use cars as an example again. Here is a potential user flow:

  • A user enters some basic criteria on what used car he/she wants to buy: the type of car, mileage, budget, etc.
  • Sellers who think they have the possible used cars this particular user likes will submit an application, almost like submitting a job application, including what his or her car is and why the buyer should pick his/hers.

You may say, isn’t it the “WANT” option on craigslist?


But there are more.

  • The website then uses algorithm to pick a few most likely to be sold cars (the number is user configurable) and then presents to the buyer.
  • The buyer can choose to pick one from the list and contact the seller directly. They are free to complete the transaction with each other offline.

You may ask: hey, where is the revenue stream?

Here it is.

  • The buyer can opt in a service that assists you to complete the transaction. The assistance could include
    • legal assistance: help finish all the paper work
    • mechanic check assistance: help inspect the car and provide accurate estimate and negotiation
  • The seller can also opt in a service that pre-inspect the selling car, which will show a proof that the car has been inspected (on what date). This can increase the likelihood that the car gets picked by a potential buyer.

I think it will work.

Unlike the bidding from eBay, which must have the sellers first. This idea serves the buyers first. The sellers must “sell”. Every seller is like a salesman on this website.

Extend Further

Let’s go further from here.

What if every E-commerce is like that?

Here is a use case.

I want to buy a book case. Traditionally, I will go to Google to find what website sells book cases. I examine all different websites for all the book cases, and read the reviews of the book case, and then read the review of the website (unless it’s a big trust-able seller).

That is totally waste of my time.

What if I enter briefly what I want in the book case, say, I want 4 levels, wood, with a stair. It should be A meter wide and B meter tall.

Now, all the sellers to submit an application of their book cases. Sure, you can argue this will only work if the sellers (human) are actively monitoring the request. But think about the possibility that we can build a robots pulling the information from all sorts of website and then intelligently select the most possible items I will be interested.

I will be given, say 10 different book cases. I am free to click the item and go to that website to complete the purchase. However, I can opt in a service that uses machine learning to tell me if those items are priced accurately. This will be useful, especially if some of those items are second-handed.


The idea is, a website serves from a buyer’s perspective, reduces the noise the buyer could receive from traditional E-commerce approach, and offer intelligent service to facilitate the buyers to find the things they most likely fall in love with.

I seriously think it might work.