Using machine learning to predict Kickstarter success

How can you optimise for success?Laura LewisBlockedUnblockFollowFollowingApr 9Project aimsIn recent years, the range of funding options for projects created by individuals and small companies has expanded considerably.

In addition to savings, bank loans, friends & family funding and other traditional options, crowdfunding has become a popular and readily available alternative.

Kickstarter, founded in 2009, is one particularly well-known and popular crowdfunding platform.

It has an all-or-nothing funding model, whereby a project is only funded if it meets its goal amount; otherwise no money is given by backers to a project.

A huge variety of factors contribute to the success or failure of a project — in general, and also on Kickstarter.

Some of these are able to be quantified or categorised, which allows for the construction of a model to attempt to predict whether a project will succeed or not.

The aim of this project is to construct such a model and also to analyse Kickstarter project data more generally, in order to help potential project creators assess whether or not Kickstarter is a good funding option for them, and what their chances of success are.

Data sourceThe dataset used in this project was downloaded in .

csv format from a webscrape conducted by a webscraping site called Web Robots.

The dataset contains data on all projects hosted on Kickstarter between the company’s launch in April 2009 until the date of the webscrape on 14 March 2019.

The dataset contains 209,222 projects, although some of these are duplicates.

Cleaning and pre-processingA fair amount of cleaning was required to get the dataset into a format suitable for applying machine learning models.

If you just can’t get enough of df.

isna() and df.

drop() you can check out the full Jupyter notebook code in my GitHub repository.

After duplicates and irrelevant rows were dropped (e.


projects which were cancelled mid-campaign, or which were still live), I was left with a decent sized dataset of 168,979 projects.

The columns which were kept or calculated were:The project goal (in USD)Campaign length — number of days from launch to deadlineNumber of days from page creation to project launchBlurb word lengthName word lengthWhether the project was highlighted as a staff pick (one-hot encoded)Category (one-hot encoded)Country (one-hot encoded)Month a project was launched in (one-hot encoded)Month of a project’s deadline (one-hot encoded)Day of the week a project was launched on (one-hot encoded)Day of the week of a project’s deadline (one-hot encoded)Two-hour time window a project was launched in (one-hot encoded)Two-hour time window of a project’s deadline (one-hot encoded)Some features were initially retained for exploratory data analysis (EDA) purposes, but were then dropped in order to use machine learning models.

These included features that are related to outcomes (e.


the amount pledged and the number of backers) rather than related to the properties of the project itself (e.


category, goal, length of campaign).

Exploratory data analysisNow for the colourful part.

Kickstarter has grown massively since its launch in 2009, particularly during 2014 when expansion really started to ramp up.

The proportion of projects that succeed decreased considerably at this point, however, as the site was flooded with a much larger number of projects.

The success rate has been on the increase in recent years though — so there’s still hope.

Overall, 56% of completed projects (i.


those that have finished and weren’t cancelled or suspended) were successful.

Changes over time in the number of projects launched on KickstarterChanges over time in project successes and failuresThe graphs below show the differences in some of the features between successful and failed projects.

The key takeaways from this are:Unsurprisingly, successful projects tend to have smaller (and therefore more realistic) goals — the median amount sought by successful projects is about half that of failed projects (medians are used due to high positive skew of funding and goal amounts).

The differences in the median amount pledged per project are more surprising.

The median amount pledged per successful project is notably higher than the median amount requested, suggesting that projects that meet their goal tend to go on to gain even more funding, and become ‘over-funded’.

On a related note, the difference between failed and successful companies is much larger in terms of amount pledged and the number of backers, compared to goal amount.

Probably once potential backers see that a project looks like it will be successful, they are much more likely to jump on the bandwagon and fund it.

Successful projects have slightly shorter campaign lengths, but take slightly longer to launch (measured from the time the project was first created on the site).

Roughly 20% of successful projects were highlighted on the site as staff picks.

It does not seem unreasonable to suggest a causative relationship here, i.


that projects that are chosen as staff picks are much more likely to go on to be successful, and that only a few staff picks go on to fail.

Comparison of features between successful and failed projectsVarious other features were explored, in terms of project number, goal and funding amounts, backers and success rates.

For example, the graphs below show the differences between different project categories (the code is also provided).

The key takeaways from this are:The best project types to launch on Kickstarter are comics (on the grounds of success rate, number of backers and amount pledged), dance (success rate and amount pledged) and games (amount pledged and number of backers).

This is probably at least partly due to their relatively small funding goals — as noted above, projects with smaller goals tend to be more successful.

Although comics and games tend to attract the most backers, each backer tends to pledge relatively little.

Dance and film & video tend to attract the most generous backers.

Technology projects have the highest median goal size by far.

However, they are towards the bottom of the leaderboard in terms of the median amount actually pledged.

The worst performing categories are food, journalism and technology.

# Code used to create the graphs below# Importing the required librariesimport pandas as pdimport matplotlib.

pyplot as pltfrom matplotlib import cmimport numpy as np# Creating a dataframe grouped by category with columns for failed and successfulcat_df = pd.





sum()# Plottingfig, ((ax1, ax2), (ax3, ax4), (ax5, ax6)) = plt.

subplots(3, 2, figsize=(12,12))color = cm.










plot(kind='bar', ax=ax1, color=color)ax1.

set_title('Number of projects')ax1.





plot(kind='bar', ax=ax2, color=color)ax2.

set_title('Median project goal ($)')ax2.





plot(kind='bar', ax=ax3, color=color)ax3.

set_title('Median pledged per project ($)')ax3.



sum(axis=1), axis=0).


plot(kind='bar', ax=ax4, color=color) # Normalizes counts across rowsax4.

set_title('Proportion of successful projects')ax4.





plot(kind='bar', ax=ax5, color=color)ax5.

set_title('Median backers per project')ax5.





plot(kind='bar', ax=ax6, color=color)ax6.

set_title('Median pledged per backer ($)')ax6.




show()Exploring the ‘category’ featureIn the interests of space and retinas, only the ‘success proportion’ graphs will be shown below for additional features.

Again, for more detail, feel free to check out my GitHub repository.

The key takeaways from this are:Hong Kong is home to a greater proportion of successful projects (they also have a higher median number of backers and amount of funding).

Tuesday is the best day to launch a project, and weekends are the worst (the same pattern holds for the amount raised and the number of backers).

12pm to 2pm UTC is the best time to launch a project — it also has the greatest median number of backers and amount of funding.

6pm to 4am UTC is the worst time to launch.

October is the best month to launch a project — it also has the greatest median number of backers and amount of funding.

July and December are the worst months.

Preparing the data for machine learningThe ultimate goal of this project was to create a model that could predict, with a good level of accuracy, whether a project was likely to succeed or fail.

In order to prepare the data for machine learning, the following steps were taken (code below):One-hot encoding categorical variables.

Separating the data into the dependent target variable ‘y’ (in this case ‘state’, i.


project success or failure) and the independent features ‘X’.

Transforming the features in X so that they are all on the same scale.

For this project, StandardScaler from Scikit-learn was used to transform each feature to a mean of 0 and a standard deviation of 1.

The data was separated into a training and test set, for robust evaluation of the models.

# Importing the required librariesimport pandas as pdfrom sklearn.

preprocessing import StandardScalerfrom sklearn.

model_selection import train_test_split# 1) Creating dummy variablesdf_transformed = pd.

get_dummies(df_transformed)# 2) Separating into X and yX_unscaled = df_transformed.

drop('state', axis=1)y = df_transformed.

state# 3) Transforming the datascaler = StandardScaler()X = pd.


fit_transform(X_unscaled), columns=list(X_unscaled.

columns))# 4) Splitting into train and test setsX_train, X_test, y_train, y_test = train_test_split(X,y, test_size=0.

3, random_state=123)It is good practice to choose an evaluation method before running machine learning models — not after.

The weighted average F1 score was chosen.

The F1 score calculates the harmonic mean between precision and recall, and is a suitable measure because there is no preference for false positives or false negatives in this case (both are equally bad).

The weighted average will be used because the classes are of slightly different sizes, and we want to be able to predict both successes and failures.

It is good practice to choose an evaluation method before running machine learning models — not after.

Model 1: vanilla logistic regressionLogistic regression can be used as a binary classifier in order to predict which of two categories a data point falls in to.

To create a baseline model to improve upon, a logistic regression model was fitted to the data, with default parameters.

# Importing the required librariesfrom sklearn.

linear_model import LogisticRegressionfrom sklearn.

metrics import classification_report# Fitting a logistic regression model with default parameterslogreg = LogisticRegression()logreg.

fit(X_train,y_train)# Making predictionsy_hat_train = logreg.

predict(X_train)y_hat_test = logreg.

predict(X_test)# Logistic regression scoresprint("Logistic regression score for training set:", round(logreg.

score(X_train, y_train),5))print("Logistic regression score for test set:", round(logreg.

score(X_test, y_test),5))print("!.Classification report:")print(classification_report(y_test, y_hat_test))Results of the vanilla linear regression modelNot bad.

The model has a weighted average F1 score of 0.


The aim is now to improve upon this score.

Principal Component AnalysisThere were a large number of features (106) in the dataset used for the initial logistic regression model.

PCA (Principal Component Analysis) was used to reduce this into a smaller number of components which still explain as much variation in the data as possible.

This can help improve model fitting and accuracy.

The graph below (produced by the code below) shows that there was no obvious cut-off for the number of components to use in PCA.

# Importing the required librariesimport matplotlib.

pyplot as pltfrom sklearn.

decomposition import PCA# Fitting PCApca = PCA()pca.

fit_transform(X)explained_var = np.


explained_variance_ratio_)# Plotting the amount of variation explained by PCA with different numbers of componentsplt.

plot(list(range(1, len(explained_var)+1)), explained_var)plt.

title('Amount of variation explained by PCA', fontsize=14)plt.

xlabel('Number of components')plt.

ylabel('Explained variance');Plotting the amount of variation in the data explained by PCA using various numbers of componentsThe following results were found:Number of components explaining 80% of variance: 58Number of components explaining 90% of variance: 70Number of components explaining 99% of variance: 90To choose the number of components to use in the machine learning models, each of these values was plugged into a pipeline for a logistic regression model using the default parameters:# Running a for loop to test different values of n_componentsn_comps = [58,70,90]for n in n_comps: pipe = Pipeline([(‘pca’, PCA(n_components=n)), (‘clf’, LogisticRegression())]) pipe.

fit(X_train, y_train) print(“.Number of components:”, n) print(“Score:”, round(pipe.

score(X_test, y_test),5))The results showed that the score is highest for 90 components, although the difference is small (c.

3% improvement from 58 components):Number of components: 58Score: 0.

67831Number of components: 70Score: 0.

6858Number of components: 90Score: 0.

70799Model 2: logistic regression with PCA and parameter optimisationThe logistic regression model can potentially be further improved by optimising its parameters.

GridSearchCV was used to test multiple different regularisation parameters (values of C), penalties (l1 or l2) and models with and without an intercept.

# Importing the required librariesfrom sklearn.

model_selection import GridSearchCV# Timing how long the model takes to runlogreg_start = time.

time()# Building the pipelinepipe_logreg = Pipeline([('pca', PCA(n_components=90)), ('clf', LogisticRegression())])# Creating the parameters to testparams_logreg = [ {'clf__penalty': ['l1', 'l2'], 'clf__fit_intercept': [True, False], 'clf__C': [0.

001, 0.

01, 1, 10] }]# Using GridSearchCV to test multiple different parametersgrid_logreg = GridSearchCV(estimator=pipe_logreg, param_grid=params_logreg, cv=5)grid_logreg.

fit(X_train, y_train)logreg_end = time.

time()logreg_best_score = grid_logreg.

best_score_logreg_best_params = grid_logreg.

best_params_# Printing the resultsprint(f"Time taken to run: {round((logreg_end – logreg_start)/60,1)} minutes")print("Best accuracy:", round(logreg_best_score,2))print("Best parameters:", logreg_best_params)Results:Time taken to run: 48.

56 minutesBest accuracy: 0.

71Best parameters: {‘clf__C’: 10, ‘clf__fit_intercept’: True, ‘clf__penalty’: ‘l2’}A classification report and a confusion matrix were then produced for the logistic regression model using the best parameters (according to the accuracy score).

The confusion matrix was produced using the following function:def plot_cf(y_true, y_pred, class_names=None, model_name=None): """Plots a confusion matrix""" cf = confusion_matrix(y_true, y_pred) plt.

imshow(cf, cmap=plt.


Blues) plt.

grid(b=None) if model_name: plt.

title("Confusion Matrix: {}".

format(model_name)) else: plt.

title("Confusion Matrix") plt.

ylabel('True Label') plt.

xlabel('Predicted Label') class_names = set(y_true) tick_marks = np.

arange(len(class_names)) if class_names: plt.

xticks(tick_marks, class_names) plt.

yticks(tick_marks, class_names) thresh = cf.

max() / 2.

for i, j in itertools.


shape[0]), range(cf.

shape[1])): plt.

text(j, i, cf[i, j], horizontalalignment='center', color='white' if cf[i, j] > thresh else 'black')plt.

colorbar()The full results for the best logistic regression model are below:Results of the best logistic regression modelAfter hyperparameter tuning, the model’s accuracy score is the same as the logistic regression model using default parameters (0.

70 weighted average F1 score).


Model 3: Random ForestsNext, a Random Forest classifier was used.

The Random Forest algorithm is a supervised learning algorithm that can be used for classification.

It works by building multiple different decision trees to predict which category a data point belongs to.

Again, GridSearchCV was used to test multiple different hyperparameters, in order to optimise the model.

# Importing the required librariesfrom sklearn.

ensemble import RandomForestClassifier# Using GridSearchCV to test multiple different parametersrf_start = time.

time()pipe_rf = Pipeline([('pca', PCA(n_components=90)), ('clf', RandomForestClassifier())])params_rf = [ {'clf__n_estimators': [100], 'clf__max_depth': [20, 30, 40], 'clf__min_samples_split':[0.

001, 0.

01] }]grid_rf = GridSearchCV(estimator=pipe_rf, param_grid=params_rf, cv=5)grid_rf.

fit(X_train, y_train)rf_end = time.

time()rf_best_score = grid_rf.

best_score_rf_best_params = grid_rf.

best_params_print(f"Time taken to run: {round((rf_end – rf_start)/60,1)} minutes")print("Best accuracy:", round(rf_best_score,2))print("Best parameters:", rf_best_params)Results:Time taken to run: 72.

2 minutesBest accuracy: 0.

7Best parameters: {'clf__max_depth': 30, 'clf__min_samples_split': 0.

001, 'clf__n_estimators': 100}The full results for the best Random Forest model are below:Results of the best Random Forest modelAfter hyperparameter tuning, the model’s weighted average F1 score increased from 0.

65 for a model with default settings to 0.


This is similar to, although slightly worse than, the logistic regression model.

Also, the difference between the score for the training set and the test set suggests there might be some over-fitting.

There may well be more scope for hyperparameter tuning here to further improve the model, but time precluded it.

Model 4: XGBoostAh, the darling of the Kaggle world.

XGBoost is very ‘in’ right now.

This is a form of gradient boosting algorithm.

Similar to Random Forests, it is an ensemble method that produces multiple decision trees to improve classification of data points, but it uses gradient descent to improve the performance of the model for the data points that are particularly difficult to classify.

Back to good ol’ GridSearchCV for hyperparameter testing:# Importing the required librariesfrom sklearn.

model_selection import GridSearchCV# Using GridSearchCV to test multiple different parametersxgb_start = time.

time()pipe_xgb = Pipeline([('pca', PCA(n_components=90)), ('clf', xgb.

XGBClassifier())])params_xgb = [ {'clf__n_estimators': [100], 'clf__max_depth': [25, 35], 'clf__learning_rate': [0.

01, 0.

1], 'clf__subsample': [0.

7, 1], 'clf__min_child_weight': [20, 100] }]grid_xgb = GridSearchCV(estimator=pipe_xgb, param_grid=params_xgb, cv=5)grid_xgb.

fit(X_train, y_train)xgb_end = time.

time()xgb_best_score = grid_xgb.

best_score_xgb_best_params = grid_xgb.

best_params_print(f"Time taken to run: {round((xgb_end – xgb_start)/60,1)} minutes")print("Best accuracy:", round(xgb_best_score,2))print("Best parameters:", xgb_best_params)Results:Time taken to run: 865.

4 minutesBest accuracy: 0.

7Best parameters: {'clf__learning_rate': 0.

1, 'clf__max_depth': 35, 'clf__min_child_weight': 100, 'clf__n_estimators': 100, 'clf__subsample': 0.


14 and a half hours, and it was still only able to achieve the same accuracy as the initial regression model (this was also only a 0.

01 increase in accuracy from an XGBoost model that was run with default parameters).

The full results for the best XGBoost model are below:Results of the best XGBoost modelAs with the Random Forest model, the difference between the accuracy score for the training set and the test set suggests there might be some over-fitting.

Again, there may well be more scope for hyperparameter tuning here to further improve the model — but I didn’t have another 14 and a half hours to spare.

Model evaluationEach model was able to achieve an accuracy of about 70%, after parameter tuning.

Although it was relatively easy to reach roughly this level of accuracy, parameter tuning was only able to increase accuracy levels by a small amount.

Possibly the reasonably large amount of data for each of only two categories meant that there was enough data for even a relatively simple model (e.


logistic regression with default settings) to achieve a good level of validation accuracy.

The best Random Forest and XGBoost models created still showed some degree of over-fitting.

Further parameter tuning would be required to attempt to reduce this.

The final chosen model is the tuned logistic regression model.

This is because, although each model was able to achieve a similar level of accuracy for the test set, this is the only model that did not exhibit overfitting.

Interestingly, each model performed worse at predicting failures compared to successes, with a lower true negative rate than true positive rate.



it classified quite a few failed projects as successes, but relatively few successful projects as failures.

Possibly the factors that might cause a project to fail are more likely to be beyond the scope of the data, e.


poor marketing, insufficient updates, or not replying to messages from potential backers.

The false positive and false negative rates mean that, if the data about a new project is fed through the model to make a prediction about its success or failure:if the project is going to end up being a success, the model will correctly predict this as a success about 80% of the timeif the project is going to end up being a failure, the model will only correctly predict this as a failure about 60% of the time (and the rest of the time will incorrectly predict it as a success)Recommendations for project creators considering KickstarterSome of the factors that had a positive effect on success rate and/or the amount of money received are:Most important:Smaller project goalsBeing chosen as a staff pick (a measure of quality)Comics, dance and games projectsProjects from Hong KongLess important:Shorter campaignsTaking longer between creation and launchFilm & video and music projects (popular categories on the site, and fairly successful)Launching on a Tuesday (although this is also the most common day to launch a project, so beware the competition)Launching in OctoberLaunching between 12pm and 2pm UTC (this is of course related to the country a project is launched from, but remember that backers can come from all over the world)Factors which had a negative effect on success rate and/or the amount of money received are:Most negative:Large goalsFood and journalism projectsProjects from ItalyLess negative:Longer campaignsLaunching on a weekendLaunching in July or DecemberLaunching between 6pm and 4am UTCOverall, Kickstarter is well suited to small, high-quality projects, particularly comics, dance and games.

It is less suited to larger projects, particularly food (e.


restaurants) and journalism projects.

Thanks for reading this far.If you have any thoughts, comments or suggestions, please add them below.


. More details

Leave a Reply