🪙 AI Model to Predict ICO Success for Fundraising Teams & Startups

Using Machine Learning Classification Algorithms | Cryptocurrency Offerings

---

📌 Business Problem

Initial Coin Offerings (ICOs) have become a popular crowdfunding mechanism for startups, but their decentralised nature creates high information asymmetry — making it difficult for investors to assess whether a project will successfully reach its fundraising goal. Fraudulent "pump and dump" ICO schemes further compound this risk.

This project builds and compares 6 Machine Learning classification models to predict the probability of ICO fundraising success, giving startups and investors a data-driven tool to evaluate ICO viability.

• Y = Successful fundraising goal achieved
• N = Unsuccessful fundraising goal

---

🏗️ Project Pipeline

ICO Dataset (2,767 projects)
        ↓
  ┌───────────────────────────────────┐
  │  Data Preprocessing               │
  │  · Missing value imputation       │
  │  · Outlier removal (Boxplot)      │
  │  · Feature encoding               │
  └───────────────────────────────────┘
        ↓
  Train / Test Split (90% / 10%)
  10-Fold Cross Validation
        ↓
  ┌──────┬──────┬──────┬──────┬─────┬──────┐
  │ KNN  │  NB  │  DT  │  RF  │ SVM │  LR  │
  └──────┴──────┴──────┴──────┴─────┴──────┘
        ↓
  Evaluation: Accuracy, AUC, Precision,
              Recall, F-measure, Kappa

---

📊 Dataset Features

Feature	Type	Description
hasVideo	Binary (0/1)	Project has a promotional video
hasReddit	Binary (0/1)	Project has a Reddit community
hasGithub	Binary (0/1)	Project has a GitHub repository
minInvestment	Binary (0/1)	Minimum investment threshold set
rating	Numeric (1–5)	Project rating score
priceUSD	Numeric	ICO token price in USD
teamSize	Numeric	Number of team members
coinNum	Numeric	Number of coins issued
distributedPercentage	Numeric	% of coins distributed
platform	Character	Platform used for ICO
success	Target (Y/N)	Whether fundraising goal was met

Key insight: Successful projects averaged a rating of 3.39, vs 2.97 for unsuccessful ones.

---

🛠️ Tech Stack

Tool	Purpose
caret	Model training & cross-validation
class	KNN classifier
tm / e1071	Naive Bayes classifier
rpart / C50	Decision Tree classifier
randomForest	Random Forest classifier
pROC / ROCR	ROC curve & AUC evaluation
ggplot2	Data visualisation
VIM	Missing value visualisation

---

📦 Setup (R)

install.packages(c("caret", "class", "e1071", "tm", "rpart", "C50",
                   "randomForest", "pROC", "ROCR", "ggplot2", "VIM",
                   "tidyr", "dplyr", "knitr", "mlbench"))

# Load dataset
icobench <- read.csv("Downloads/Machinelearningcourseworkright.csv")
str(icobench)
summary(icobench)

---

🔬 Key Steps

1. Missing Value Imputation

library(VIM)
# Impute missing values with column mean
si_icobench$priceUSD[is.na(si_icobench$priceUSD)] <- mean(si_icobench$priceUSD, na.rm = TRUE)
si_icobench$teamSize[is.na(si_icobench$teamSize)] <- mean(si_icobench$teamSize, na.rm = TRUE)

2. Outlier Removal

# Outliers found in: priceUSD, teamSize, coinNum, distributedPercentage
outlier <- boxplot(data$priceUSD, plot = FALSE)$out
x <- data[-which(data$priceUSD %in% outlier), ]

3. KNN Model (Best Performer)

library(caret)
K <- 47
knn <- train(success ~ ., data = train, method = "knn",
             preProcess = c("center", "scale"),
             trControl = trainControl(method = "repeatedcv",
                                      number = 10, repeats = 3,
                                      classProbs = TRUE),
             metric = "ROC", tuneLength = 10)

4. ROC & AUC Evaluation

library(pROC)
roc <- roc(test$success, test_pred)
plot(roc, print.auc = TRUE, auc.polygon = TRUE,
     grid = c(1,2), grid.col = c("green","red"),
     max.auc.polygon = TRUE, auc.polygon.col = "lightblue")
auc(roc)

---

📈 Model Performance Results

Classifier	Accuracy (%)	AUC (%)	Precision	Recall	F-Measure	Kappa	Category
KNN	100	100	1.00	1.00	1.00	1.00	🥇 Ultimate
Decision Tree	81	61	0.86	0.83	0.85	0.60	🥈 Highline
SVM	71	70	0.72	0.90	0.80	0.30	🥈 Highline
Random Forest	69	65	0.83	0.74	0.78	0.26	Baseline
Logistic Regression	66	67	0.67	0.84	0.75	0.32	Baseline
Naive Bayes	63	70.5	0.98	0.65	0.77	0.12	Baseline

---

💡 Key Findings

• KNN achieved 100% accuracy (CCI: 100%, ICI: 0%, AUC: 1.000) — outperforming all other models across every metric by incorporating all dataset attributes including priceUSD and teamSize
• DT and SVM are Highline classifiers — solid performers in the 70–85% accuracy range, suitable for simpler use cases
• NB, RF, and LR are Baseline classifiers — limited because they do not fully leverage numeric attributes to define fundraising success
• Rating threshold matters — projects with ratings ≥ 3.3 show significantly higher success probability (visible in Decision Tree splits)
• Proper pre-processing is critical — outliers in priceUSD, teamSize, coinNum, and distributedPercentage significantly impact model accuracy if not removed

---

🗂️ Repository Structure

├── data/
│   └── icobench.csv              # ICO project dataset (2,767 records)
├── models/
│   ├── preprocessing.R           # Missing values & outlier handling
│   ├── knn_model.R               # KNN classifier
│   ├── naive_bayes.R             # Naive Bayes classifier
│   ├── decision_tree.R           # Decision Tree (C5.0)
│   ├── random_forest.R           # Random Forest
│   ├── svm_model.R               # Support Vector Machine
│   └── logistic_regression.R     # Logistic Regression
├── outputs/
│   ├── roc_curves/               # ROC plots for all 6 models
│   └── confusion_matrices/       # Confusion matrix outputs
└── README.md

---

⚠️ Limitations

• KNN's perfect accuracy may reflect overfitting on this specific dataset — further validation on unseen ICO data is recommended
• Models do not incorporate whitepaper quality, social media sentiment, or market conditions at time of ICO launch
• False negatives carry high business risk — a fundraising team predicted as "unsuccessful" may still succeed with external factors

---

👤 Author

Aniket Amar Nerali MSc Business Analytics & Decision Sciences — University of Leeds