How Designers’ Irrationality & Inertia Make Up Sribu.com’s Success

Sribu is a website that connects between client who need graphic designs and community of designers from all around the world (taken from sribu.com). It’s a business modelled after 99design in the US. While 99 is an international success, I was really skeptical when Sribu started locally. In my eyes, our people are not fond of working for free. Why would they work for free if websites like freelancer.co.id and the likes could guarantee you’d get the money after doing a design project? If I had a billion dollars, I wouldn’t invest a penny in Sribu; To say I was skeptical was an understatement. But it turns out I was completely wrong. The model seems to work just as well in Indonesia as validated by fundings they’ve obtained so far.

So in my journey to study data science, I decided to analyse Sribu

Any marketplace faces at least 2 opposing markets: people who want to buy and people who provides the things to buy. Unless they were born out of money, founders starting up a marketplace always face a classic chicken and egg problem: do I grow the buyers first or do I grow the sellers first? Whichever path they take, they eventually need to show that their marketplace is alive—that there are people in need of (design) services and there are people who provide (design) services.

Sribu’s success depends on the number of successful contests. We can use number of design submitted as a proxy of a contest success (and in turn, Sribu’s success). I define a contest as a success if there are sufficient amount of design submissions. What qualifies as sufficient is outside the scope of the study. (Fun fact: when I conducted this study, there are 4 contests receiving no entries at all).

Then the question of money becomes: what are the factors that determine a contest’s success? If the team knew what are the driving factors for success, they can optimise their efforts to increase these factors.

Starting from the most sensible factor: rewards

Each contest in Sribu has a monetary prize up for grabs. In an efficient market and if their designers were completely rational, a contest’s prize has a direct relationship with a contest’s submission—the bigger the reward, the more attractive it becomes.

But apparently, being sensible with the prize amount is not a sufficient prerequisite of a design marketplace success. As we can see in the chart (data truncated to only include those with more than 500 submissions), there are contests offering very minimal prize with a lot of submissions. Vice versa, there are some contests with large reward but the number of submissions is not that much. That is, designers on Sribu are not completely rational.

If they were, we would be seeing something like this instead

So it has to be something more than just money

Using machine learning, this is what I found.

First of all, what this chart shows: on the y axis, we have a list of features that matter for any design project success on Sribu.com. To clarify some terms used:

• brief: the length of the project brief (number of words)
• extend: there is an extension for the project due date
• confidential: submitted designs will not be shown to public even after the contest is closed.
• private: submitted designs will not be shown to public while the contest is open

On the x axis, we have the magnitude of impact towards # of entries for a given feature. For example, keeping everything else the same, a project in logo design will have about 126 more submissions, on average, than projects in other categories. Another example, projects on designing calendar will generate about 40 less submissions, on average.

It turns out money is not the most important factor. Neither how long and descriptive the brief is. The type of work for the contest matters the most. Common type of requests such as Logo, Stationery, T-shirt design generate a lot of participations compared to a more particular request such as booth design or even web design. In other words, the market is inefficient; It does not produce at the maximum capacity. It’s like having a machine capable of doing 15 things but it is only regularly used to do 3.

What does it mean for Sribu.com?

I’m not sure how this trend will persist but human behaviors tend to follow the law of inertia: unless there is an external force exerted on a moving object, its trajectory will persist. Although the inertia is one of the startup’s achilles' heels, it has also been the driver of its success as we’ve seen in the above analysis. The current state of market inefficiency can be turned into an opportunity for effort concentration.

And that has a lot of implication for user acquisition on service demand side: Sribu can grow the platform significantly and organically just by hosting more “typical” design contests. Find out which companies need these kind of service the most and focus on getting them onboard.

The team should also give some thought on how to scale the variety of designers. Logo and stationery are not things that change often. However, banner, invitation, booth, flyer designs are marketing collaterals that have high turnover. Repeatable and significant monetization comes from these kinds of work. The challenge for the product and marketing team is how to increase designer interest on these type of work. The demand is clearly there: flyer design is the 5th most requested, while web design is sitting on rank #7 (4.2% and 2.7% compared to total project, respectively; Logo design occupies 50% of the whole demand)

Both prize and the length of contest brief are significant predictors of the number of entries, but their contribution magnitude for success is much smaller than the type of work.

Disclaimer: I sincerely hope this analysis can be of some help for Sribu team. Feel free to use any findings and conclusions. But please double-check with your own analysis to be extra careful.

Research process:

• data collection: I scrape Sribu.com to collect the data.
• feature selection: I simply grab the informations publicly available. I assume that, by and large, designers make their judgment based on description for the design work. Features in the model have been further selected using backward selection based on significance value.
• data pre-processing:
• filling in prize manually for a USD 2000 contest. For some reason the prize turned out to be NA
• as there are many private/confidential contests requiring logins, the briefs could not be obtained by simple scraping which result in NA. I imputed these NA values by first grouping all contests into 1 of 5 clusters using kmeans algorithm (k=5). If a contest has NA for brief value, assign the mean brief value of the cluster to which the contest belongs.
• I removed an outlier in an attempt to get a better linear fit.
• My train/test split resulted in my test set having an observation for Mobile Apps Design that is not contained by my training set. Thus, I removed the observation from my test set.
• software used: I only use R for this analysis.
• R packages involved: XML, dplyr, ggplot2
• algorithms involved: linear regression (using base lm package), kmeans

Notes:

• Prize stated in this project is in USD
• I split my training and test set 50:50 using seed(123)
• Project R code is going to be uploaded in my github repo after some serious cleaning (it was really messy..). UPDATE: it's up!
• This project was conducted in the second week of February. If you source the code, your mileage may vary
• The number of available projects when I conducted this study was 2292, which is not a lot. As such, my training set contains some data points with small number of instances for some features. If you use a different seed for splitting, your mileage may vary
• One could argue that since I only use linear regression for descriptive and not predictive analysis, I should’ve just used all the data in my model instead of splitting between train and test set. The reason I did it is because I want to test the predictive ability of the model to make sure it holds a reasonable degree of generalizability.

Originally posted in my tumblr.