Quant Research Guide
Firm-Specific Preparation
Generic prep is not enough. Each top quant firm has a distinct culture, research philosophy, and interview format. The difference between a generic candidate and one who has done firm-specific research is visible immediately. Know each firm's unique focus, understand their published work, and tailor your signal research story to fit their model.
Core Theory
- 1Two Sigma: a technology and data science company that happens to trade. Research is highly collaborative, cross-functional, and data-driven. Unique emphasis on ML/AI, alternative data, and systematic processes at every step. Values candidates who can bridge quantitative finance and data science. Their interview emphasizes practical problem-solving: given messy data, what do you do?
- 2D.E. Shaw: intellectually eclectic, academically rigorous, and deliberately broad in strategy. One of the most selective interview processes in finance โ they interview fewer people but more thoroughly. Values mathematical depth and creative thinking. Unusual fact: D.E. Shaw himself (founder) has a computer science background, not finance โ which shaped the culture toward algorithmic thinking and formal reasoning.
- 3WorldQuant: an 'alpha factory' model. Rather than a few large bets, they pursue hundreds of small, uncorrelated alpha signals across global equities, futures, and other markets. Their interview focuses heavily on signal writing using their WebSim platform โ you write alpha expressions in a specific format and they evaluate IC, Sharpe, and drawdown immediately. Volume and diversity of alphas is the key metric.
- 4Citadel Equities: a fundamental long-short fund with quantitative overlay. Different from purely systematic quant funds โ sector specialists apply fundamental views which quantitative tools help express. Requires understanding both factor models AND sector-specific fundamental analysis. Their interview often includes sector deep dives alongside quant methodology questions.
- 5Point72 Asset Management: Steve Cohen's multi-manager platform with both discretionary fundamental and quantitative strategies. The quant pod model allows individual portfolio managers significant autonomy. Values candidates who can work independently, generate ideas end-to-end, and communicate research clearly to PMs with varying quantitative backgrounds.
- 6Millennium Management: the original multi-PM model. Hundreds of independent portfolio managers each allocated capital based on risk-adjusted performance. Quantitative PMs here must show alpha generation track record or compelling research. Interviews focus heavily on your personal track record and specific signal ideas โ 'what would you trade on day one?'
- 7AQR Capital Management: the academic quant house. Founded by Cliff Asness (Chicago Booth PhD), known for rigorous published research and factor investing (value, momentum, carry, defensive). AQR publishes most of their research publicly โ read their papers before interviewing. They value academic-quality research and theoretical depth above most other firms.
- 8Jane Street: primarily an options market maker and ETF arbitrageur, not a fundamental quant fund. Interviews focus heavily on probability puzzles, expected value calculations, options pricing intuition, and game theory. If you're targeting Jane Street, probability puzzle preparation is even more critical than signal research.
Key Formulas
| 1 | import numpy as np |
| 2 | import pandas as pd |
| 3 | from scipy.stats import spearmanr |
| 4 | |
| 5 | np.random.seed(42) |
| 6 | T, N = 252, 100 # 252 days, 100 stocks |
| 7 | tickers = [f'STK{i:03d}' for i in range(N)] |
| 8 | |
| 9 | # Simulate OHLCV data |
| 10 | close = pd.DataFrame(100 * np.exp(np.cumsum(np.random.randn(T, N)*0.01, axis=0)), columns=tickers) |
| 11 | volume = pd.DataFrame(np.random.lognormal(10, 0.5, (T, N)), columns=tickers) |
| 12 | returns = close.pct_change() |
| 13 | |
| 14 | # โโโ Alpha 1: Short-term Reversal โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ |
| 15 | # WorldQuant expression: -rank(ts_return(close, 5)) |
| 16 | # Stocks with biggest 5-day losses are expected to reverse |
| 17 | def ts_return(df, d): return df.pct_change(d) |
| 18 | def rank_cs(df): return df.rank(axis=1, pct=True) |
| 19 | |
| 20 | alpha1 = -rank_cs(ts_return(close, 5)) |
| 21 | |
| 22 | # โโโ Alpha 2: Momentum โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ |
| 23 | # WorldQuant: rank(ts_return(close, 252)) - rank(ts_return(close, 21)) |
| 24 | # 12m momentum minus 1m return (12-1 style) |
| 25 | alpha2 = rank_cs(ts_return(close, 252)) - rank_cs(ts_return(close, 21)) |
| 26 | alpha2 = rank_cs(alpha2) # re-rank the combination |
| 27 | |
| 28 | # โโโ Alpha 3: Volume-Price Interaction โโโโโโโโโโโโโโโโโโโโโโโโโโโโ |
| 29 | # Stocks with high volume on down days = bearish signal |
| 30 | # -rank(correlation(returns, volume, lookback=10)) |
| 31 | def rolling_corr(df1, df2, window): |
| 32 | result = pd.DataFrame(index=df1.index, columns=df1.columns) |
| 33 | for col in df1.columns: |
| 34 | result[col] = df1[col].rolling(window).corr(df2[col]) |
| 35 | return result.astype(float) |
| 36 | |
| 37 | # Simplified cross-sectional corr each day |
| 38 | price_vol_corr = rolling_corr(returns, np.log(volume), window=10) |
| 39 | alpha3 = -rank_cs(price_vol_corr) |
| 40 | |
| 41 | # โโโ Evaluate IC for each alpha โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ |
| 42 | fwd_ret = returns.shift(-1) # next-day return |
| 43 | |
| 44 | def compute_ic(alpha, fwd): |
| 45 | alpha_s = alpha.shift(1) # look-ahead prevention |
| 46 | ics = [] |
| 47 | for t in range(len(alpha_s)): |
| 48 | if alpha_s.iloc[t].notna().sum() > 20: |
| 49 | ic, _ = spearmanr(alpha_s.iloc[t].dropna(), |
| 50 | fwd.iloc[t].reindex(alpha_s.iloc[t].dropna().index)) |
| 51 | ics.append(ic) |
| 52 | ic_arr = np.array([x for x in ics if not np.isnan(x)]) |
| 53 | return ic_arr.mean(), ic_arr.std(), ic_arr.mean()/ic_arr.std() |
| 54 | |
| 55 | print(f"{'Alpha':<10} | {'Mean IC':>8} | {'IC Std':>8} | {'ICIR':>8}") |
| 56 | print("-" * 42) |
| 57 | for name, alpha in [('Reversal', alpha1), ('Momentum', alpha2), ('Vol-Price', alpha3)]: |
| 58 | ic_mean, ic_std, icir = compute_ic(alpha, fwd_ret) |
| 59 | print(f"{name:<10} | {ic_mean:>8.4f} | {ic_std:>8.4f} | {icir:>8.4f}") |
| 60 | |
| 61 | # โโโ Alpha combination (WorldQuant style) โโโโโโโโโโโโโโโโโโโโโโโโโ |
| 62 | combined = rank_cs(alpha1 + alpha2.fillna(0)) |
| 63 | ic_c, _, icir_c = compute_ic(combined, fwd_ret) |
| 64 | print(f"{'Combined':<10} | {ic_c:>8.4f} | {' ---':>8} | {icir_c:>8.4f}") |
Implements three WorldQuant-style alpha expressions in Python: short-term reversal (-rank of 5-day return), 12-1 momentum combination, and a volume-price correlation signal. Each alpha is cross-sectionally ranked, combined, and evaluated by IC and ICIR โ exactly the workflow used on the WorldQuant Brain/WebSim platform.
Worked Interview Problems
3 problemsProblem
Two Sigma gives you 10 years of daily stock returns plus 5 features. They say: 'Explore this data. What do you find?' How do you structure your 45-minute response?
Solution
Minutes 0-5: Data QA and overview. 'First I'd check for missing data, outliers, and distributional properties. I'd compute summary statistics for each feature and the return distribution.' State what you find concisely.
Minutes 5-15: Univariate IC analysis. 'I'd compute the cross-sectional IC between each feature and next-day/next-month returns. I'd plot IC over time to identify which features are predictive and whether the relationship is stable.'
Minutes 15-25: Multivariate exploration. 'I'd check pairwise correlations between features to understand redundancy, then build a combined signal using IC-weighted combination or a simple Ridge regression. I'd evaluate OOS IC with walk-forward.'
Minutes 25-35: Drill into the most interesting finding. 'Feature 3 has IC = 0.05 at 1-week horizon, decaying to 0 by month 2. This looks like a short-term reversal pattern. I'd check whether it's driven by liquidity (small/micro-cap stocks) or is universal.'
Minutes 35-45: Summarize and discuss extensions. 'The strongest signal is Feature 3 at 1-week horizon, IC = 0.05, ICIR = 0.7. Next steps: test after costs (high turnover signal), check for sector concentration, combine with Feature 1 which has low correlation.'
Answer
Structure: QA (5min) โ univariate IC (10min) โ multivariate (10min) โ deep dive on best finding (10min) โ summary + extensions (10min). Think out loud throughout โ Two Sigma evaluates process as much as conclusion.
Common Mistakes
Giving the same generic answer about 'why this firm' to every company. Interviewers at Two Sigma, D.E. Shaw, and WorldQuant can immediately tell when a candidate hasn't read anything about the firm. Read their published research, blog posts, or investor letters before every interview.
Not practicing WorldQuant alpha expression writing before a WorldQuant interview. The WebSim platform has a specific notation (ts_mean, ts_arg_max, rank, delay, etc.). If you've never written an alpha expression in their format, you will struggle under interview pressure. Spend 10+ hours on Brain/WebSim before interviewing.
Underestimating the math depth at D.E. Shaw. Their interviews are among the most mathematically rigorous in the industry โ expect to derive estimators, prove properties of distributions, and work through probability problems at graduate level. Generic interview prep is not sufficient.
Not reading AQR's published research before an AQR interview. Cliff Asness and the AQR team publish extensively โ 'Fact, Fiction, and Momentum Investing', 'Value and Momentum Everywhere', 'The Devil in HML's Details'. Reading 2-3 of their papers signals genuine interest and gives you specific talking points.
Framing the wrong research style for each firm. At Citadel Equities, fundamental understanding of sectors matters alongside quant methodology. At WorldQuant, diversity and volume of alphas matters more than one strong signal. At D.E. Shaw, the quality of mathematical reasoning matters most. Tailor your research story accordingly.