Machine Learning in Fintech-Enabled Investment Analytics

Data and preliminary analysis (SPY test dataset)

1. Introduction

1.1 Background Financial innovation has changed how money moves, and with it how illicit money moves. Mobile money platforms have widened financial access across Sub-Saharan Africa and other emerging regions, but the same high-volume, low-value transaction streams that make them useful also give launderers room to hide. Nahimana and Gaba (2026) describe this tension in Rwanda, where several million users transact through telecom-led wallets such as MTN and Airtel, and where the Financial Intelligence Centre must supervise transaction flows whose scale and speed exceed what static, rule-based monitoring can handle. Detection is only half of the problem. An alert has no justice value unless a legal system can turn it into evidence, a prosecution, and recovered assets. FATF (2021) finds that environmental crime, including forestry crime, is highly profitable and that offenders use front companies to mix legal and illegal goods and payments early in resource supply chains. Whether such flows can be traced depends on legal and institutional design, and in the United States and the European Union that design is now moving in different directions: the EU’s AML Regulation applies from 10 July 2027, while a final FinCEN rule of August 2026 removed domestic companies from US beneficial-ownership reporting. These two literatures rarely speak to each other. One is technical and asks how to find suspicious flows at scale. The other is legal and asks how to act on them. This project sits at the seam between the two. 1.2 Problem Statement Machine-learning monitoring systems are usually judged by detection metrics such as precision, recall, and alert volume. Legal enforcement systems are judged by whether evidence is admissible, traceable, and sufficient to link proceeds to a predicate crime. There is no established way to translate the first kind of output into the second. A model score is not evidence, and an alert is not a case. As a result, even a well-performing detection pipeline may produce little enforcement value if its outputs cannot support a follow-the-money investigation. 1.3 Research Gap Nahimana and Gaba (2026) deliberately stop at operational alerting and outline a path toward validation with regulators. The enforcement sources, including FATF’s (2021) typology work and Sitompul’s (2026) doctrinal study of Indonesia, describe how funds from environmental crime are traced and prosecuted but do not test detection models. As far as I can tell from the sources reviewed so far, none examines what lies between an ML-generated alert and a prosecutable follow-the-money case in the United States or the European Union, or whether behavioral features built for generic laundering typologies also capture laundering linked to resource-extraction crimes such as illegal logging. 1.4 Research Questions 1. What does a machine-learning alert need to contain, and which downstream steps must follow, to support a follow-the-money investigation of laundering from environmental crime in the United States and in the European Union? 2. How well do account-centric behavioral features (velocity, net-flow direction, counterparty diversity, burstiness) separate laundering typologies that resemble resource-crime proceeds from other typologies? 3. How do US and EU rules on predicate offences, beneficial-ownership transparency, and institutional coordination differ, and what do those differences mean for sharing alert data between financial institutions, financial intelligence units, and environmental or customs authorities? 1.5 Objectives and Significance The objective is to build a conceptual and empirical bridge between detection and enforcement: first by mapping model outputs to legal evidentiary needs, then by testing, on public synthetic data, whether detection features carry typology-specific signal that investigators could use. The project matters for three reasons. It connects the fintech and analytics side of AML to the rule-of-law side. It compares how the United States and the European Union, two systems moving in different directions on beneficial-ownership transparency, would need to receive and act on ML-generated alerts. And it offers design ideas for regulators and financial institutions that must turn alerts into cases. 1.6 Structure of the Paper Section 2 reviews the technical and legal literatures. Section 3 sets out the proposed methodology and data analysis plan. Later sections (to be drafted) will present results, discussion, and conclusions.

2. Literature Review

This review rests on the two core papers, the FATF report on laundering from environmental crime, and primary and secondary sources on US and EU law, current as of October 2026. Legal descriptions come largely from law-firm and institutional summaries and must be checked against primary texts before submission. 2.1 Machine Learning for Transaction Monitoring in Mobile Money Rule-based monitoring remains the operational default in AML because it is auditable and easy to explain to supervisors. Nahimana and Gaba (2026) note, citing the wider literature, that fixed thresholds generate large numbers of false positives, miss temporal and network structures typical of structuring and layering, and adapt slowly to new typologies. Machine learning is positioned as a complement: it learns nonlinear feature interactions from historical activity and ranks alerts by risk rather than by rule violation. Their study works under three constraints that define real AML practice: extreme class imbalance (about 0.1 percent laundering prevalence), scarce and delayed labels because confirmed suspicious transaction reports arrive months after the event, and a hard ceiling on investigator capacity. Using SAML-D, a synthetic dataset of 9,504,852 transactions covering 17 laundering typologies, they engineer causal, account-centric behavioral features (rolling velocity, net-flow directionality, counterparty diversity, burstiness) under strict leakage controls and a chronological train, validation, and test split. They benchmark supervised classifiers (logistic regression, random forest, LightGBM), unsupervised anomaly detectors (isolation forest, local outlier factor), a dense autoencoder, and a logistic-regression late-fusion meta-learner. The evaluation is operational rather than accuracy-based, which is appropriate given the imbalance. LightGBM reaches a PR-AUC of 0.0469 and captures 64 true laundering cases at precision of about 0.89 with 0.51 alerts per 10,000 transactions. The fusion model reaches a PR-AUC of 0.0477 at precision of about 0.91 and 0.46 alerts per 10,000, but recovers 59 true positives, so fusion trims alert volume slightly without adding recall. The authors map score bands to analyst workflows and suspicious transaction report escalation, and they state that their contribution is a governance-aware pipeline and evaluation protocol, not a new algorithm. Two limits matter for this project: the data are synthetic, and the pipeline ends at alert prioritization. 2.2 Follow the Money and Enforcement: International, US, and EU Frameworks The enforcement sources fall into three layers: the international standard, US law, and EU law. Indonesia is kept as a single comparator. International standard. FATF (2021) describes environmental crime, including forestry crime, illegal mining, and waste trafficking, as a low risk, high reward crime. Its report draws on case studies from over 40 countries and shows criminals using front companies to mix legal and illegal goods and payments early in supply chains. Its handout for environmental-crime authorities urges them to set up communication channels and share information proactively with financial investigators and AML authorities, and notes that money laundering laws are often more dissuasive than environmental laws. United States. US money laundering law is built around specified unlawful activities (SUAs) set by statute, and liability reaches not only offenders but also bankers, accountants, and lawyers who facilitate the transactions (CFA Institute, 2018). For illegal logging, the main trade-side tool is the Lacey Act, which since its 2008 amendments prohibits importing illegally harvested wood and carries forfeiture, fines, and imprisonment (Gregg & Porges). Secondary sources say a Lacey Act violation may also trigger smuggling or money laundering charges, while one civil-society source says Lacey Act violations are not themselves predicate offences. I will settle this against the text of 18 U.S.C. 1956(c)(7) before submission. On beneficial ownership, the Corporate Transparency Act of 2021 created a federal reporting regime, but FinCEN’s final rule of 11 August 2026, effective 14 August 2026, exempts domestic companies and US persons from reporting. Foreign companies registered to do business in the US must still report their non-US beneficial owners. My reading, to be checked, is that this leaves customer due diligence at financial institutions as the main source of domestic beneficial-ownership information. European Union. The EU is moving toward a single rulebook. Regulation (EU) 2024/1624 (AMLR) applies from 10 July 2027, tightens customer due diligence, and harmonises beneficial-ownership rules around ownership or control, with 25 percent or more of shares or voting rights as the ownership indicator. Regulation (EU) 2024/1620 created the Anti-Money Laundering Authority (AMLA) in Frankfurt, scheduled to start operating on 1 July 2025, and Directive (EU) 2024/1640 covers financial intelligence units, supervision, and beneficial-ownership registers. Most obliged entities stay under national supervision, with AMLA coordinating and directly supervising selected cases. On criminal law, Directive (EU) 2018/1673 lists 22 predicate-offence categories, including environmental crime. The revised Environmental Crime Directive (EU) 2024/1203 adds offences, penalties for natural and legal persons, and stronger cross-border cooperation. Member States had to transpose it by 21 May 2026, and Eurojust published a practitioners’ note around that date. I have not yet checked which Member States met the deadline. Comparator: Indonesia. Sitompul (2026) analyses Indonesia’s 2010 money laundering law and 2023 Criminal Code, which keep forestry crime as a predicate offence. Following Ginting (2021) and Firdaus and Arifin (2021), he treats follow the money as tracing asset flows rather than only pursuing offenders. He reports that the approach is held back by institutional silos, opaque beneficial ownership, and weak coordination among agencies. I use these three obstacles as a checklist when examining the US and EU, not as a separate jurisdiction of study. The paper is normative and does not test data. 2.3 Synthesis and Gap The sources complement each other. Nahimana and Gaba address detection in a data-rich, label-scarce setting and supply an evaluation language (alerts per 10,000, recall at fixed precision). FATF and the US and EU instruments describe what enforcement requires once funds are identified, and Sitompul shows where coordination breaks down. Side by side, they expose four open questions: • Evidentiary translation. Detection outputs are risk scores, while prosecutions need traceable, attributable evidence tied to a predicate offence. • Typology specificity. Generic behavioral features do not appear to have been examined for fit to resource-crime laundering, such as mixing legal and illegal payments early in supply chains. • Coordination as a data problem. Gaps between financial, environmental, and customs authorities are partly gaps in shared data and access. • Diverging transparency regimes. The EU is harmonising and verifying beneficial-ownership data, while the US has removed domestic reporting, which changes what investigators can trace. 2.4 Sources to Add Before Submission Before submission, the review should add: the EU Deforestation Regulation and timber due diligence rules; US case law on the Lacey Act and the SUA list; US suspicious activity reporting and customer due diligence rules; Member State transposition of Directive 2024/1203; FATF guidance on beneficial ownership; and graph and network methods for AML. These are a to-do, not reviewed material. Legal summaries here need checking against the Official Journal, the U.S. Code, and the Federal Register.

3. Data

The preliminary analysis uses daily prices of the SPDR S&P 500 ETF (SPY) from Yahoo Finance, 2015 to 2025. The question is whether ML models predict next-day returns better than a simple benchmark out of sample (H1), and whether any advantage survives trading costs (H2).

3.1 Load the data

library(quantmod)
library(randomForest)
library(knitr)

START    <- as.Date("2015-01-01")
END      <- as.Date("2025-12-31")
SPLIT    <- as.Date("2022-01-01")  # train before, test from this date
COST_BPS <- 5                      # assumed cost per unit of trading, in basis points

load_prices <- function() {
  if (file.exists("spy.csv")) {
    df <- read.csv("spy.csv")
    df$Date <- as.Date(df$Date)
    return(df)
  }
  spy <- getSymbols("SPY", src = "yahoo", from = START, to = END,
                    auto.adjust = TRUE, auto.assign = FALSE)
  df <- data.frame(Date = index(spy),
                   Close = as.numeric(Cl(spy)),
                   Volume = as.numeric(Vo(spy)))
  df <- na.omit(df)
  write.csv(df, "spy.csv", row.names = FALSE)
  df
}

raw <- load_prices()
cat(nrow(raw), "trading days,", format(min(raw$Date)), "to", format(max(raw$Date)), "\n")

2765 trading days, 2015-01-02 to 2025-12-30 

kable(tail(raw, 3))

	Date	Close	Volume
2763	2025-12-26	690.31	41613300
2764	2025-12-29	687.85	62559500
2765	2025-12-30	687.01	47160700

3.2 Features and target

Every feature uses information up to the previous day only, so the model never sees the day it predicts.

lagk       <- function(x, k = 1) c(rep(NA, k), head(x, -k))
pct_change <- function(x, k = 1) c(rep(NA, k), x[-(1:k)] / head(x, -k) - 1)
roll_sd    <- function(x, w) sapply(seq_along(x), function(i)
                if (i < w) NA else sd(x[(i - w + 1):i]))

ret <- pct_change(raw$Close, 1)

data <- data.frame(
  Date    = raw$Date,
  target  = ret,
  ret_1d  = lagk(ret, 1),
  ret_5d  = lagk(pct_change(raw$Close, 5), 1),
  ret_21d = lagk(pct_change(raw$Close, 21), 1),
  vol_21d = lagk(roll_sd(ret, 21), 1),
  dvolume = lagk(pct_change(raw$Volume, 1), 1)
)
data$dvolume[is.infinite(data$dvolume)] <- NA
data <- na.omit(data)

features <- c("ret_1d", "ret_5d", "ret_21d", "vol_21d", "dvolume")
train <- data[data$Date <  SPLIT, ]
test  <- data[data$Date >= SPLIT, ]
cat("Train:", nrow(train), "days | Test:", nrow(test), "days\n")

Train: 1741 days | Test: 1002 days

stats <- t(sapply(data[, c("target", features)],
  function(x) c(mean = mean(x), sd = sd(x), min = min(x), max = max(x))))
kable(stats, digits = 4)

	mean	sd	min	max
target	0.0005	0.0113	-0.1094	0.1050
ret_1d	0.0005	0.0113	-0.1094	0.1050
ret_5d	0.0025	0.0230	-0.1797	0.1736
ret_21d	0.0104	0.0448	-0.3314	0.2519
vol_21d	0.0094	0.0063	0.0022	0.0588
dvolume	0.0519	0.3546	-0.6876	2.3600

3.3 Models and out-of-sample accuracy (H1)

The benchmark predicts the training-period mean return every day. The out-of-sample R-squared compares each model’s squared errors with that benchmark, so the benchmark scores 0 by construction and a positive value means the model adds information.

y_te <- test$target
frm  <- as.formula(paste("target ~", paste(features, collapse = " + ")))

set.seed(42)
ols <- lm(frm, data = train)
rf  <- randomForest(frm, data = train, ntree = 300, nodesize = 20)

preds <- list(
  "Benchmark (train mean)" = rep(mean(train$target), nrow(test)),
  "OLS"                    = as.numeric(predict(ols, test)),
  "Random forest"          = as.numeric(predict(rf, test))
)

sse_bench <- sum((y_te - preds[[1]])^2)
accuracy <- do.call(rbind, lapply(names(preds), function(nm) {
  p <- preds[[nm]]
  data.frame(Model = nm,
             RMSE = sqrt(mean((y_te - p)^2)),
             OOS_R2 = 1 - sum((y_te - p)^2) / sse_bench,
             Directional_accuracy = mean(sign(p) == sign(y_te)))
}))
kable(accuracy, digits = 4)

Model	RMSE	OOS_R2	Directional_accuracy
Benchmark (train mean)	0.0114	0.0000	0.5369
OLS	0.0115	-0.0244	0.5140
Random forest	0.0115	-0.0204	0.5000

3.4 Trading rule and costs (H2)

Each model drives a simple long/short rule: hold SPY when the forecast is positive and short it when negative. Costs are charged on every change of position (a switch from long to short counts as two units of trading).

run_rule <- function(pred, actual, cost_bps) {
  pos <- ifelse(pred >= 0, 1, -1)
  turnover <- c(abs(pos[1]), abs(diff(pos)))
  pos * actual - turnover * cost_bps / 10000
}
sharpe <- function(x) sqrt(252) * mean(x) / sd(x)

models <- c("OLS", "Random forest")
curves <- list("Buy and hold" = y_te)
summary_tab <- data.frame()
for (nm in models) {
  gross <- run_rule(preds[[nm]], y_te, 0)
  net   <- run_rule(preds[[nm]], y_te, COST_BPS)
  curves[[paste(nm, "(net)")]] <- net
  summary_tab <- rbind(summary_tab, data.frame(
    Strategy = nm,
    Gross_Sharpe = sharpe(gross),
    Net_Sharpe = sharpe(net),
    Net_total_return = prod(1 + net) - 1))
}
summary_tab <- rbind(summary_tab, data.frame(
  Strategy = "Buy and hold", Gross_Sharpe = sharpe(y_te),
  Net_Sharpe = sharpe(y_te), Net_total_return = prod(1 + y_te) - 1))
names(summary_tab)[3] <- paste0("Net_Sharpe_", COST_BPS, "bps")
kable(summary_tab, digits = 3)

Strategy	Gross_Sharpe	Net_Sharpe_5bps	Net_total_return
OLS	0.505	-0.151	-0.159
Random forest	-0.046	-0.667	-0.420
Buy and hold	0.605	0.605	0.446

growth <- sapply(curves, function(r) cumprod(1 + r))
matplot(test$Date, growth, type = "l", lty = 1, lwd = 2,
        col = c("black", "steelblue", "firebrick"),
        xlab = "", ylab = "Growth of 1 unit")
legend("topleft", legend = names(curves), lty = 1, lwd = 2,
       col = c("black", "steelblue", "firebrick"), bty = "n")

Figure 1: Cumulative return in the test period (net of assumed costs)

How sensitive is the result to the cost assumption?

sens <- do.call(rbind, lapply(c(0, 2, 5, 10), function(b) {
  data.frame(Cost_bps = b,
             OLS_net_Sharpe = sharpe(run_rule(preds[["OLS"]], y_te, b)),
             RF_net_Sharpe  = sharpe(run_rule(preds[["Random forest"]], y_te, b)))
}))
kable(sens, digits = 3)

Cost_bps	OLS_net_Sharpe	RF_net_Sharpe
0	0.505	-0.046
2	0.242	-0.295
5	-0.151	-0.667
10	-0.801	-1.282

3.5 Optional: dictionary sentiment score

To add the sentiment strand, download the Loughran-McDonald master dictionary CSV from the Notre Dame SRAF site and score headlines or sentences. This chunk is switched off until the file is in place.

lm_dict <- read.csv("Loughran-McDonald_MasterDictionary.csv")
pos_words <- tolower(lm_dict$Word[lm_dict$Positive > 0])
neg_words <- tolower(lm_dict$Word[lm_dict$Negative > 0])

lm_score <- function(text) {
  words <- strsplit(tolower(gsub("[^a-z' ]", " ", tolower(text))), "\\s+")[[1]]
  words <- words[words != ""]
  (sum(words %in% pos_words) - sum(words %in% neg_words)) / max(length(words), 1)
}
lm_score("Profits fell sharply and the outlook worsened.")

3.6 Interpretation

(Write this after the first real run. Answer in your own words, with the numbers from the tables above.)

• H1: Does either model have a positive out-of-sample R-squared? Is directional accuracy meaningfully above 50%?
• H2: Does the net Sharpe ratio fall below the gross one and below buy and hold? At what cost level does the advantage disappear?
• Limits: one ETF, one test period, no transaction-size or short-selling costs, no tuning of hyperparameters.

References

(Paste from your doc.)