Company Interview Questions · Banking & Fintech

Goldman Sachs & Morgan Stanley: Engineering (ML) Interviews

❓ 6 questions🏢 Goldman Sachs & Morgan Stanley🆕 Recent reports

Goldman Sachs & Morgan Stanley: Engineering Analyst (ML) and Tech Interviews (2025). Rounds, real reported questions and model answers, compiled from public candidate reports (sources at the end). Practice mode hides every answer: say your answer first, press Show answer, then rate yourself. Use 🔁 Review again to build your revision list and 🎲 Random question for a quick drill.

★ Must-know 145📝 Assignments
About this interview

Compiled from public candidate reports: Goldman Sachs Engineering Analyst with an ML-focused round (Bengaluru, June 2025, waitlisted) and other Bengaluru reports (2025), plus Morgan Stanley technology intern reports (Bengaluru and Mumbai, 2025) that show the coding bar. The questions are paraphrased, and the model answers are written by EasyLearnSpace.

Goldman Sachs Morgan Stanley
Role Engineering Analyst (ML team); Engineering Analyst Technology / SDE intern; GenAI developer roles listed
Rounds R1: ML + DSA (explain and code parts of your model; HashMap, in-order/post-order traversal, sorting); R2: one coding problem (in-order variant). Others: OA (DP, graphs) → 3 technical/HR rounds OA (SHL) → technical → HR; or technical → culture fit
Reported questions Retrieve values from a dataset, ML terminology, BST insert, print the tree boundary Quicksort complexity, linked lists, OS basics, design Uber
Outcome Waitlisted, then not selected; others not selected Not selected (interns)

01Rounds at a glance

02Questions and model answers

Q1
Print the boundary (outer layer) of a binary tree anticlockwise.
def boundary(root):
    if not root: return []
    is_leaf = lambda n: not n.left and not n.right
    res = [root.val] if not is_leaf(root) else []
    n = root.left                                   # left boundary (top-down, no leaves)
    while n:
        if not is_leaf(n): res.append(n.val)
        n = n.left or n.right
    def leaves(n):
        if not n: return
        if is_leaf(n): res.append(n.val); return
        leaves(n.left); leaves(n.right)
    leaves(root)
    right, n = [], root.right                       # right boundary (bottom-up, no leaves)
    while n:
        if not is_leaf(n): right.append(n.val)
        n = n.right or n.left
    return res + right[::-1]
Q2
Insert a value into a binary search tree.
def insert(root, val):
    if not root: return Node(val)
    if val < root.val: root.left = insert(root.left, val)
    else: root.right = insert(root.right, val)
    return root
  • O(h) time: O(log n) when balanced, O(n) when skewed. Mention self-balancing trees (AVL, red-black).
Q3
Explain your ML model and write its core training code.
  • Framing: "Default prediction with gradient boosting. Features: payment history, utilisation, tenure. Target: 90+ days past due within 12 months."
  • Code skeleton:
from sklearn.model_selection import train_test_split
from sklearn.metrics import roc_auc_score
import lightgbm as lgb
Xtr, Xte, ytr, yte = train_test_split(X, y, test_size=.2, stratify=y, random_state=0)
m = lgb.LGBMClassifier(n_estimators=500, learning_rate=.05, num_leaves=31, class_weight="balanced")
m.fit(Xtr, ytr, eval_set=[(Xte, yte)], callbacks=[lgb.early_stopping(50)])
print("AUC", roc_auc_score(yte, m.predict_proba(Xte)[:, 1]))
  • Be ready to explain every parameter, the validation choice and the feature importance.
Q4
Quicksort: what is its time complexity, and how do you avoid the worst case?
  • Average O(n log n), worst case O(n²) when the pivots are consistently the smallest or largest (for example, sorted input with a first-element pivot).
  • Avoid it: random pivot, median-of-three, introsort (switch to heapsort at a recursion-depth limit), and three-way partitioning for many duplicates.
  • It sorts in place (O(log n) stack) and is not stable.
Q5
High-level design: design a ride-hailing system like Uber.
  • Core services:
  • Rider and driver apps.
  • Location service (drivers send GPS every ~4 s, stored in a geo-index: geohash or quadtree, in Redis).
  • Matching service (find nearby available drivers, ranked by ETA).
  • Trip service (state machine).
  • Pricing (surge = f(demand/supply per zone)).
  • Payments.
  • Notifications.
  • Data: a trips DB (SQL for consistency), location data (in-memory plus a time-series store), Kafka for events.
  • Scale: partition by city or region, WebSockets for live updates, idempotent payment APIs.
  • ML hooks: ETA prediction, surge forecasting, fraud detection, driver–rider matching optimisation.
Q6
ML terminology quick-fire: precision, recall, AUC, regularisation, cross-validation, feature leakage.
  • Precision = TP/(TP+FP). Recall = TP/(TP+FN).
  • AUC = the probability that a random positive ranks above a random negative.
  • Regularisation penalises complexity (L1/L2) to reduce overfitting.
  • Cross-validation rotates validation folds for a robust estimate.
  • Leakage means a feature contains information unavailable at prediction time, which inflates offline scores.
  • Give one-line definitions plus an example each; quick-fire rounds reward crispness.

03How to prepare

Sources

Created by Ajeetkumar