Master Outline & Study Guide

Python Financial Analysis — Portfolio Performance & Risk (2015–2024)

Purpose of this document: A complete, self-contained reference for studying this portfolio project. Use it to prepare for interviews — to explain what was built, how it works, why each decision was made, and to answer probing technical questions. When studying with Claude, this is the file Claude reads to teach and quiz you.

One sentence to anchor everything: This project pulls 10 years of daily market data for six diversified assets via the yfinance API and runs it through a five-notebook Python pipeline — EDA → returns & risk → correlation & diversification → time-series decomposition → ARIMA forecasting — to evaluate portfolio performance, quantify risk, and demonstrate the value of diversification.

Important — read §13 before an interview. The portfolio write-up (index.md) describes a few techniques (an ADF stationarity test, ARIMA confidence intervals, covariance-matrix diversification math, VaR, SARIMA, ML regression) that are not in the committed notebooks. This guide is written from what the notebooks actually do, and §13 lays out the gaps so nothing surprises you.

Table of Contents

1. The 30-Second Pitch
2. Why This Project Exists (Context)
3. The Tech Stack
4. The Dataset — 6 Assets via yfinance
5. The 5-Notebook Pipeline (Architecture)
6. Notebook 1 — Data Acquisition & EDA
7. Notebook 2 — Returns & Risk Analysis
8. Notebook 3 — Correlation & Diversification
9. Notebook 4 — Time Series Decomposition & Trends
10. Notebook 5 — ARIMA Forecasting
11. Key Results & Numbers
12. Finance & Stats Concepts to Know Cold
13. Limitations & Honest Caveats (incl. write-up vs. notebooks)
14. Design Decisions & Trade-offs (the “Why”)
15. Interview Q&A
16. How to Walk Through This Project Live
17. Glossary

1. The 30-Second Pitch

This project is an end-to-end Python financial analysis of a six-asset diversified portfolio over 10 years of daily data (2015–2024). It retrieves the data programmatically via the yfinance API, then works through five Jupyter notebooks: (1) data acquisition and exploratory analysis, (2) returns and risk metrics, (3) correlation and diversification, (4) time-series decomposition and trends, and (5) an ARIMA forecasting baseline.

The six assets span asset classes deliberately — a market benchmark (SPY), a growth stock (AAPL), an energy-sector ETF (XLE), long-term bonds (TLT), gold (GLD), and international equities (EFA) — so the analysis can study real cross-asset relationships. The pipeline computes risk-adjusted performance (Sharpe ratio, maximum drawdown, annualized volatility), maps correlations, demonstrates that a diversified allocation cuts volatility, decomposes the S&P 500 into trend/seasonal/residual components, and fits an ARIMA model to forecast SPY prices.

The headline findings: enormous performance dispersion (AAPL +936%, TLT −11% over the decade), diversification meaningfully reduced volatility, equity assets are too correlated to diversify each other, and forecasting price levels is genuinely hard — the ARIMA baseline captured trend direction but missed regime shifts.

One-line version: “I built a five-notebook Python pipeline that pulls 10 years of market data via an API and runs the full quant-analyst workflow — risk metrics, correlation and diversification analysis, time-series decomposition, and ARIMA forecasting — on a six-asset diversified portfolio.”

2. Why This Project Exists (Context)

The premise. Investment decisions are too often driven by intuition — “stocks go up, bonds are safe.” This project’s job is to replace intuition with evidence: take real market data and quantify performance, risk, and the actual benefit of diversification.

The simulated role. The project plays a data analyst supporting portfolio management and investment decision-making — turning raw price data into the risk and return diagnostics a portfolio manager needs.

Why these six assets. They were chosen to be a genuinely diversified portfolio across asset classes, so the analysis exercises real cross-asset behavior — equities vs. bonds vs. gold vs. international — rather than six tech stocks that all move together. That choice is what makes the correlation and diversification analysis (Notebook 3) meaningful.

Why it’s a strong portfolio project. It demonstrates the full quantitative-analysis loop on real financial data: API data engineering (programmatic retrieval, validation), risk analytics (Sharpe, drawdown, volatility), statistical analysis (correlation, time-series decomposition), and predictive modeling (ARIMA) — all in Python, all reproducible. It’s a finance-flavored data-science project that shows breadth across data engineering, statistics, and modeling.

The honest framing. The project is deliberately structured as a learning pipeline — five notebooks that build on each other, each ending in a completion summary. It’s exploratory and descriptive analysis, not a production trading system, and the forecasting notebook is explicit about that (“Forecasting is HARD! Don’t use for real trading!”).

3. The Tech Stack

The mental model. yfinance is the data source, pandas/NumPy are the analysis engine, matplotlib/seaborn are the presentation layer, and statsmodels/scikit-learn are the modeling layer. Everything runs on a laptop in seconds; no GPU, no cloud.

One architectural fact to know. The notebooks hand off through CSV files. Notebook 1 saves cleaned_prices_2015_2024.csv; Notebooks 2–5 load it (and Notebook 2 saves daily_returns and cumulative_returns CSVs for Notebook 3). So the pipeline is a chain: each notebook reads the previous notebook’s saved output. (The data files committed in data/processed/ carry an .xls extension but are actually plain CSVs — a minor mislabeling.)

4. The Dataset — 6 Assets via yfinance

No static data file ships with the project — all price data is retrieved programmatically via the yfinance API (Yahoo Finance), for January 2015 through December 2024.

The six assets — and why each one:

The data. Daily price data, auto_adjust=True (so prices are adjusted for splits and dividends — important for accurate return math). The analysis uses the Close column. The result: 2,515 trading days per asset — roughly 15,000 daily price points across the six assets, with zero missing values (a 100%-complete dataset, validated in Notebook 1).

Why auto_adjust=True matters (a likely interview question): adjusted close prices back out the artificial price drops caused by stock splits and dividend payouts, so a computed return reflects an investor’s actual total return rather than a cosmetic price change. Using unadjusted prices would understate returns for dividend-paying assets.

The committed processed files (data/processed/): cleaned_prices_2015_2024 (2,515 rows — the daily close prices), daily_returns_2015_2024 (2,514 rows — one fewer, because pct_change drops the first day), and cumulative_returns_2015_2024 (2,514 rows — the growth-of-$1 series).

5. The 5-Notebook Pipeline (Architecture)

  yfinance API  (Yahoo Finance — 6 tickers, 2015–2024 daily)
        │
        ▼
  NOTEBOOK 1 — Data Acquisition & EDA
  download (retry logic) · validate · price history chart · summary stats · correlation
        │  saves → cleaned_prices_2015_2024.csv
        ▼
  NOTEBOOK 2 — Returns & Risk Analysis
  daily & cumulative returns · volatility · Sharpe · max drawdown · risk-return scatter
        │  saves → daily_returns.csv, cumulative_returns.csv
        ▼
  NOTEBOOK 3 — Correlation & Diversification
  correlation matrix · rolling 252-day correlations · 100% SPY vs 60/30/10 portfolio
        │
        ▼
  NOTEBOOK 4 — Time Series Decomposition & Trends
  50/200-day moving averages · multiplicative seasonal decomposition
        │
        ▼
  NOTEBOOK 5 — ARIMA Forecasting
  80/20 chronological train/test split · ARIMA(1,1,1) · MAE / RMSE / MAPE
        │
        ▼
  12 saved chart PNGs in outputs/figures/

The key architectural idea — sequential, file-chained notebooks. Each notebook is one analysis phase, and they pass data forward through saved CSV files: Notebook 1 produces the clean prices, Notebook 2 produces the returns, and Notebooks 3–5 load whichever they need. The chain means each notebook is independently runnable (given the upstream CSVs exist) and the analysis builds in layers — raw data → returns → relationships → structure → forecast.

A style note. The notebooks are all code cells (no markdown cells) — narration is done through print() statements and section-header comments. Each notebook ends with a printed “completion summary” listing what it did and what it found. It’s a tutorial-style structure: clear and linear.

6. Notebook 1 — Data Acquisition & EDA

File: 01_data_acquisition_eda.ipynb · Output: cleaned_prices_2015_2024.csv, 2 charts

Business question. What are the fundamental characteristics of each asset’s 10-year price history, and what preliminary relationships exist between them?

What the notebook actually does, step by step:

1. Download with retry logic. yf.download() for the six tickers, 2015–2024, auto_adjust=True, wrapped in a 3-attempt retry loop with a 2-second wait between tries — defensive coding against a flaky API.
2. Extract Close prices and validate quality. Pulls data['Close'], checks shape and date range, and runs a missing-data check — confirmed zero missing values across all 2,515 days × 6 assets.
3. 10-year price history chart — all six assets’ raw prices on one plot.
4. Summary statistics — describe(), plus three computed performance metrics:
   • Total return = (last price / first price − 1) × 100.
   • Annualized return (CAGR) = (last/first)^(1/years) − 1 — the constant yearly growth rate that would produce the observed total return.
   • Annualized volatility = daily_returns.std() × √252 × 100.
5. Correlation matrix + heatmap — an initial prices.corr() heatmap, and it programmatically finds the highest and lowest correlations.
6. Save cleaned_prices_2015_2024.csv for the downstream notebooks.

Key findings.

• Massive performance dispersion — AAPL +935.85% total return over the decade, SPY +240.81%, and TLT −11.29% (long bonds lost money in a rising-rate decade).
• 100% data completeness — zero missing values; the yfinance API delivered a clean, reliable dataset.
• Highest volatility: XLE (29.81%); lowest: GLD (14.12%).

The teachable point. This notebook’s job is foundation: get the data, prove it’s clean, and establish the basic shape (who grew, who’s volatile, who’s correlated). A core discipline shown here: validate data quality before any downstream analysis — a missing value undetected at this stage would corrupt every return and risk calculation that follows.

7. Notebook 2 — Returns & Risk Analysis

File: 02_returns_risk_analysis.ipynb · Output: daily_returns + cumulative_returns CSVs, 3 charts

Business question. Which assets delivered the best risk-adjusted returns, and where were the largest downside episodes?

What the notebook does:

1. Daily returns — prices.pct_change().dropna() — the day-over-day percentage change.
2. Cumulative returns — (1 + daily_returns).cumprod() — the “growth of $1”: what one dollar invested in 2015 became.
3. Annualized volatility — daily_returns.std() × √252. (The √252 converts a daily standard deviation to an annual one — 252 is the standard trading-days-per-year.)
4. Annualized return (CAGR) and the Sharpe ratio — (annual return − risk-free rate) / annualized volatility, using a 2% risk-free rate.
5. Risk-return scatter plot — each asset plotted as volatility (x) vs. return (y), with diagonal lines marking Sharpe-ratio levels; the upper-left quadrant is best.
6. Maximum drawdown — the worst peak-to-trough decline, computed by tracking the running maximum (cummax) of the cumulative-return path and measuring the largest drop below it.
7. Save the daily and cumulative returns CSVs.

Key findings (approximate annualized figures over 2015–2024, Sharpe at rf = 2%):

The headline reads: AAPL had the best return and the best Sharpe; XLE had AAPL- level volatility but tiny return and a brutal −62% drawdown (worst risk-adjusted performer); TLT lost money and still drew down ~48%; GLD had the smallest drawdown (~21%) — the steadiest diversifier.

The teachable point — risk has two faces. Volatility and drawdown tell different stories: TLT’s volatility (~16%) looked moderate, but its drawdown (~48%) was severe. Volatility measures day-to-day wobble; drawdown measures the worst sustained loss an investor would actually have to live through. The lesson the notebook draws: evaluate Sharpe and drawdown together — neither alone is enough.

8. Notebook 3 — Correlation & Diversification

File: 03_correlation_diversification.ipynb · Output: 3 charts

Business question. How stable are cross-asset correlations over time, and how much risk can diversification actually remove?

What the notebook does:

1. Full-period correlation matrix on daily returns + an enhanced heatmap; it programmatically identifies the strongest and weakest pairs.
2. Rolling 252-day correlations — for four key pairs (SPY-TLT, SPY-GLD, AAPL-SPY, XLE-TLT), a one-trading-year rolling window tracks how each relationship shifts across calm and crisis periods, with the COVID window shaded.
3. The diversification benefit — it builds two portfolios as weighted daily-return series and compares them: a concentrated portfolio (100% SPY) vs. a diversified portfolio (60% SPY / 30% TLT / 10% GLD), computing each one’s annualized return, volatility, Sharpe, and drawdown.

Key findings.

• The strongest correlation is EFA–SPY at ≈ 0.855 — international and US large-cap equities move almost in lockstep. (Note: the portfolio write-up’s approximate table shows SPY-AAPL as the top pair at ~0.80; the notebook’s actual computed output says EFA-SPY 0.855 is the strongest — cite the notebook’s number.)
• Equity assets cluster — SPY, AAPL, EFA, XLE all correlate moderately-to-strongly, so adding another equity barely diversifies anything.
• TLT (bonds) is negatively correlated with stocks (~−0.35 vs. SPY) — a genuine diversifier; GLD (gold) is near-zero correlated with both stocks and bonds — an independent return driver.
• Diversification worked — the 60/30/10 portfolio had lower volatility than 100% SPY (the notebook reports a volatility reduction of ~6.67 percentage points; the write-up frames the same result as roughly a 35% relative reduction). The diversified portfolio had a better Sharpe ratio despite a lower raw return — the project’s clearest demonstration of diversification as a “free lunch.”
• Correlations are regime-dependent — the rolling windows showed the SPY-TLT negative correlation weakening (even flipping positive) during the 2022 rate-hike period, and equity correlations spiking together during the March 2020 crash — diversification weakens exactly when you need it most.

The teachable point. This is the project’s core finance lesson: diversification means combining assets with low correlation, not just holding more assets. Six equity ETFs that all correlate at 0.8 are not a diversified portfolio. And because correlations move, diversification has to be monitored, not set once.

9. Notebook 4 — Time Series Decomposition & Trends

File: 04_time_series_trends.ipynb · Output: 2 charts

Business question. What components drive SPY’s price movement, and can trend, seasonality, and noise be separated?

What the notebook does:

1. Moving averages — computes SPY’s 50-day and 200-day moving averages and overlays them on the price. These are the classic technical-analysis regime indicators: when the 50-day crosses above the 200-day it’s a “golden cross” (bullish); crossing below is a “death cross” (bearish).
2. Seasonal decomposition — statsmodels’ seasonal_decompose with a multiplicative model and a 252-day period (one trading year), splitting SPY into three components:
   • Trend — the long-run direction.
   • Seasonal — the repeating annual pattern.
   • Residual — what’s left: random noise and shocks.

Why a multiplicative model (a likely interview question): financial prices grow roughly exponentially (a percentage each year), so the components combine by multiplication (trend × seasonal × residual), not addition. An additive model fits series that grow by a fixed amount each period; a multiplicative model fits series that grow by a fixed rate — which is how markets behave.

Key findings.

• Trend dominates. The notebook’s own estimate: SPY’s behavior is roughly 80% trend, 15% seasonal, 5% random — a persistent long-run uptrend (from ~$180 to ~$520+) broken by sharp structural breaks (the March 2020 COVID crash, the 2022 rate selloff).
• Seasonality is weak — calendar patterns contribute far less than trend; they’re not a reliable forecasting signal for equities.
• The residual concentrates the shocks — crisis events show up as large residual spikes, which the trend view hides.
• Moving-average crossovers confirmed regime changes (multiple golden crosses, a brief COVID death cross) but lag actual reversals.

The teachable point. Decomposition is a diagnostic — it tells you the series is trend-driven, which is useful context before forecasting. (Note: the portfolio write-up describes an Augmented Dickey-Fuller stationarity test in this notebook; the committed 04_time_series_trends.ipynb does not include one — see §13. The standard finance fact to know anyway: raw prices are non-stationary but daily returns are stationary, which is why return-based inputs are the right choice for modeling.)

10. Notebook 5 — ARIMA Forecasting

File: 05_forecasting.ipynb · Output: 2 charts

Business question. Can a classical ARIMA model produce useful out-of-sample SPY price forecasts?

What the notebook does:

1. Chronological 80/20 train/test split — the first 2,012 days (≈2015–2022) train the model; the last 503 days (≈2022–2024) are the holdout test. Chronological is essential: you must never train on future data and test on the past — that’s lookahead leakage. The split is visualized.
2. Fit ARIMA(1,1,1) on the training SPY price series. The (p,d,q) order means: p=1 autoregressive term (today depends on yesterday), d=1 differencing (difference the series once to make it stationary), q=1 moving-average term (today depends on yesterday’s forecast error).
3. Forecast the full ~503-day test horizon with fitted_model.forecast() and plot predictions vs. actuals.
4. Evaluate with three error metrics: MAE (mean absolute error — average dollar error), RMSE (root mean squared error — penalizes large errors more heavily), and MAPE (mean absolute percentage error).

Key findings.

• ARIMA(1,1,1) captured the general upward trend direction but could not anticipate the magnitude or timing of the 2023–2024 rally — the model struggled where it mattered.
• RMSE exceeded MAE — that gap means a few large errors (during volatile windows) drove a disproportionate share of total error.
• The honest conclusion, stated in the notebook itself: “Stock prices are inherently difficult to forecast — past performance ≠ future results,” and “don’t use for real trading.” ARIMA(1,1,1) is a transparent baseline, not a production model.

The teachable point — and the maturity signal. The most valuable thing in this notebook is that it doesn’t oversell the model. A naive project would report the error metric and call the forecast a success. This one frames ARIMA as a baseline benchmark — the simple, interpretable model that any fancier approach must beat — and is candid that univariate linear models can’t predict nonlinear market regime shifts. Being able to say “my forecast was a deliberately simple baseline, and here’s why it has limits” is stronger than claiming a stock predictor that works.

11. Key Results & Numbers

Memorize the headline figures (these are from the notebooks’ actual computed output).

The five takeaways (the story the project tells):

1. Performance dispersion is enormous — over the same decade, AAPL grew ~10× and TLT lost money. Asset class is destiny.
2. Return and risk are linked — the biggest growers (AAPL) carried the biggest interim risk; the steady assets (GLD) grew modestly.
3. Path risk matters as much as endpoint return — XLE and TLT had drawdowns (~62%, ~48%) that real investors would struggle to hold through.
4. Diversification works — but only across low-correlated assets. Mixing equities, bonds, and gold cut volatility; mixing equities with more equities would not.
5. Forecasting price levels is hard — the ARIMA baseline got direction but missed magnitude and timing.

12. Finance & Stats Concepts to Know Cold

A finance/data interview will probe the fundamentals behind the project.

Daily return — (today's price / yesterday's price) − 1; in pandas, pct_change().

Cumulative return — the compounded growth of $1: (1 + daily_returns).cumprod().

Total return vs. annualized return (CAGR) — total return is the whole-period change; CAGR is the constant yearly rate that produces it: (end/start)^(1/years) − 1. CAGR makes assets held for different periods comparable.

Volatility — the standard deviation of returns; the standard risk measure. Annualized volatility = daily std × √252 — the “√time” rule scales a daily figure to a yearly one (252 = trading days per year).

Sharpe ratio — (return − risk-free rate) / volatility — risk-adjusted return, i.e., how much excess return you earn per unit of risk. Higher is better. This project uses a 2% risk-free rate. A negative Sharpe (like TLT’s) means the asset underperformed cash.

Maximum drawdown — the largest peak-to-trough percentage decline over the period; measures the worst loss an investor would have lived through. Computed by comparing the cumulative-return path to its running maximum (cummax).

Correlation — how two assets move together, from −1 (perfect opposite) to +1 (perfect together); 0 = independent. The basis of diversification.

Diversification — combining low-correlated assets so the portfolio’s volatility is less than the weighted average of its parts. Called a “free lunch” because it cuts risk without proportionally cutting return.

Risk-free rate — the return on a “riskless” asset (a Treasury bill); the baseline the Sharpe ratio measures excess return above.

Stationarity — a series whose statistical properties (mean, variance) don’t change over time. Prices are non-stationary (they trend); returns are stationary. Stationarity is a prerequisite for ARIMA — handled by the differencing step.

Time-series decomposition — splitting a series into trend (long-run direction), seasonal (repeating cycle), and residual (noise). Multiplicative model for exponentially-growing series (markets); additive for linearly-growing ones.

Moving average — the rolling mean over a window (50-day, 200-day); a smoothed trend line. The golden cross (50 crosses above 200) and death cross (below) are regime signals.

ARIMA(p,d,q) — AutoRegressive Integrated Moving Average — a classic univariate forecasting model. p = autoregressive lags, d = differencing order (to reach stationarity), q = moving-average lags. This project uses ARIMA(1,1,1).

Train/test split (chronological) — for time series, the test set must be the latest data and the train set the earliest — never random — to avoid training on the future.

MAE / RMSE / MAPE — forecast error metrics. MAE = average absolute error; RMSE = root-mean-squared error (penalizes large misses more); MAPE = average absolute percentage error. RMSE > MAE signals a few big errors.

13. Limitations & Honest Caveats (incl. write-up vs. notebooks)

Volunteer these — and especially know the first one, because it’s about your own materials.

13.1 — The portfolio write-up describes more than the committed notebooks contain. This is the most important caveat for an interview. The project’s index.md page and README.md are polished and, in several places, describe a more sophisticated version of the analysis than the five notebooks actually implement. Specifically:

• ADF stationarity test — index.md (Analysis 4) shows adfuller() code and an “ADF Statistic < −20, p-value < 0.001” result. Notebook 4 does not import or run adfuller — it has moving averages and seasonal_decompose only.
• ARIMA confidence intervals — index.md (Analysis 5) shows get_forecast() / conf_int() and says “confidence intervals were included.” Notebook 5 uses plain .forecast() and computes/plots no confidence intervals.
• Covariance-matrix diversification math — index.md (Analysis 3) shows w.T @ cov @ w portfolio-variance code. Notebook 3 builds the weighted-return series directly and never uses a covariance matrix.
• Value at Risk (VaR) — listed in index.md’s “Skills” and the README’s features. No VaR calculation appears in any notebook.
• SARIMA and machine-learning regression — the README mentions “ARIMA/SARIMA” and a “Machine Learning” notebook with “regression and classification.” Notebook 5 fits only ARIMA(1,1,1) — no SARIMA, no ML model.

How to handle this in an interview: describe what the notebooks actually do — they’re solid and real. If asked about VaR, SARIMA, ADF, or confidence intervals, be honest: “the write-up frames those as part of the analysis, but in the committed notebooks I implemented [the actual thing]; those are natural extensions I’d add next.” Never claim a technique the notebooks don’t contain. (The cleanest fix, if you want the materials to align, is to actually add those pieces to the notebooks — but as it stands, cite the notebooks.)

13.2 — Other genuine limitations:

• Backward-looking, single 10-year window. All conclusions describe 2015–2024 — a decade of mostly-rising stocks and (until 2022) low rates. A strong Sharpe here doesn’t predict the next decade.
• It’s descriptive, not a strategy. The project analyzes performance and risk; it doesn’t build or backtest a trading strategy, and the diversified 60/30/10 weights are illustrative, not optimized (no efficient-frontier / mean-variance optimization).
• ARIMA is a weak forecaster for prices — by the project’s own admission. A univariate linear model can’t capture market regime shifts.
• Survivorship and asset-selection bias — AAPL was chosen knowing it was a huge winner; picking a single growth stock in hindsight isn’t a neutral test.
• The notebooks have no markdown cells — narration is via print() statements. They read more like scripts than documented analysis notebooks.
• The processed data files are mislabeled .xls but are actually CSVs.

14. Design Decisions & Trade-offs (the “Why”)

Interviewers reward “why” answers. The deliberate choices and their rationale:

Why retrieve data via an API instead of a static CSV? Reproducibility and freshness — anyone can re-run the pipeline and get current data, and there’s no large data file to commit. The trade-off is a dependency on the API being up, which is why Notebook 1 wraps the download in retry logic.

Why auto_adjust=True? So prices are adjusted for splits and dividends. Returns computed from unadjusted prices would show a fake drop on every dividend date and split — adjusted prices give the investor’s true total return.

Why six assets across asset classes, not six stocks? The whole point is to study cross-asset relationships and diversification. Six tech stocks would all correlate near 1.0 and there’d be nothing to learn about diversification — the deliberate spread (equity, bond, gold, international) is what makes Notebook 3 meaningful.

Why annualize with √252? 252 is the standard number of US trading days per year. Volatility scales with the square root of time, so a daily standard deviation becomes annual by multiplying by √252 — this puts every asset on a comparable yearly scale.

Why compute both Sharpe and drawdown? They measure different risks. Sharpe rewards smooth return-per-unit-of-wobble; drawdown measures the worst sustained loss. TLT proved why you need both — moderate volatility but a severe drawdown.

Why a chronological train/test split for the ARIMA model? Time-series forecasting must respect time. A random split would let the model train on 2024 data and “predict” 2018 — lookahead leakage that makes the forecast look far better than it is. The chronological 80/20 split simulates genuine real-world forecasting.

Why ARIMA(1,1,1) specifically? It’s the simplest non-trivial ARIMA — one AR term, one differencing, one MA term. The point wasn’t to find the best forecaster; it was to establish a transparent, interpretable baseline that any more complex model should be benchmarked against.

Why a five-notebook pipeline instead of one big notebook? Each notebook is one phase with a clear input and output, chained through saved CSVs. It keeps each stage focused, independently runnable, and easy to follow — the analysis builds in layers.

15. Interview Q&A

Practice these out loud. Answers are written the way you’d actually speak.

Q1. Give me the overview of this project. “It’s an end-to-end Python financial analysis of a six-asset diversified portfolio over ten years of daily data, 2015 to 2024. I pull the data via the yfinance API, then work through five notebooks: data acquisition and EDA, returns and risk metrics, correlation and diversification, time-series decomposition, and an ARIMA forecasting baseline. The assets span asset classes on purpose — S&P 500, Apple, energy, long bonds, gold, international — so I can study real cross-asset behavior. The big themes are quantifying risk-adjusted return and showing what diversification actually buys you.”

Q2. How did you get the data, and how did you make sure it was clean? “Programmatically, through the yfinance API — no static file. I wrapped the download in a three-attempt retry loop in case the API hiccuped. I used auto_adjust so prices account for splits and dividends. Then before any analysis I ran a data-quality check — the dataset was 100% complete, zero missing values across 2,515 trading days for all six assets. Validating completeness first matters, because a missing value undetected there corrupts every return and risk number downstream.”

Q3. What’s the Sharpe ratio and why does it matter? “It’s risk-adjusted return — excess return over the risk-free rate, divided by volatility. It answers ‘how much return am I getting per unit of risk I’m taking?’ I used a 2% risk-free rate. It matters because raw return is misleading — XLE and AAPL both had about 30% volatility, but AAPL turned that risk into roughly a 0.80 Sharpe while XLE was near zero. Same risk, completely different reward. TLT actually had a negative Sharpe — it underperformed cash.”

Q4. What’s the difference between volatility and maximum drawdown? “Volatility is the standard deviation of returns — day-to-day wobble. Maximum drawdown is the worst peak-to-trough decline — the deepest sustained loss an investor would actually have to sit through. They can disagree: TLT had only moderate volatility, around 16%, but a roughly 48% drawdown. Volatility says ‘bumpy,’ drawdown says ‘how bad does it get and stay.’ You need both — I’d never screen a portfolio on volatility alone.”

Q5. How did you demonstrate the value of diversification? “In notebook three I compared two portfolios as weighted return series — a concentrated one, 100% SPY, and a diversified one, 60% SPY, 30% TLT, 10% gold. The diversified portfolio had clearly lower volatility — and a better Sharpe ratio even though its raw return was lower. That’s the ‘free lunch’ of diversification: combining low-correlated assets reduces risk more than it reduces return. The key is low-correlated — bonds and gold, because they don’t move with stocks.”

Q6. Why doesn’t holding more stocks diversify a portfolio? “Because equities share the same underlying risk factors. In my data, SPY, Apple, international stocks, and energy all correlated moderately to strongly — the strongest pair, EFA and SPY, was about 0.855. When assets correlate near one, they fall together, so adding another one barely reduces risk. Real diversification needs assets that move differently — that’s why I added bonds and gold, not a seventh stock.”

Q7. You found correlations change over time — why does that matter? “I computed rolling 252-day correlations, and the relationships weren’t stable. The stock-bond correlation, normally negative, weakened and even turned positive during the 2022 rate-hike period — so the bond hedge wasn’t working when stocks fell. And in the March 2020 crash, equity correlations spiked together. The lesson: diversification tends to break down exactly during a crisis, when you need it most. So correlation has to be monitored as a live risk metric, not assumed.”

Q8. Walk me through the time-series decomposition. “I decomposed SPY with a multiplicative seasonal model and a 252-day period, splitting it into trend, seasonal, and residual components. Multiplicative because markets grow roughly exponentially — by a percentage, not a fixed dollar amount — so the components multiply. The takeaway was that SPY is overwhelmingly trend-driven, roughly 80% trend, with weak seasonality and the residual concentrating the crisis shocks. That told me a forecasting model should focus on trend, and that seasonality isn’t a reliable equity signal.”

Q9. Explain your ARIMA forecasting setup. “I forecasted SPY prices with ARIMA(1,1,1) — one autoregressive term, one differencing step, one moving-average term. I used a chronological 80/20 train/test split — about 2,000 days to train, 500 to test — chronological so the model never trains on the future. Then I evaluated with MAE, RMSE, and MAPE. The honest result: the model captured the general upward direction but missed the magnitude and timing of the 2023–24 rally. RMSE exceeded MAE, meaning a few big errors during volatile periods dominated. I treated ARIMA as a transparent baseline, not a real predictor.”

Q10. Why is forecasting stock prices so hard, and is your model useful? “Stock prices are close to a random walk with nonlinear regime shifts, and a univariate linear model like ARIMA can’t anticipate a regime change — a rate shock, a crisis. So no, I wouldn’t trade on it, and the notebook says so explicitly. But it’s still useful as a baseline: it’s simple and interpretable, and any more sophisticated model — multivariate, regime-switching, machine learning — should have to beat this benchmark to justify its complexity. Knowing the limits of your model is part of the analysis.”

Q11. Why a chronological split instead of a random train/test split? “Because it’s time-series data. A random split would let the model see 2024 data during training and then ‘predict’ 2018 — that’s lookahead leakage, and it makes the forecast look artificially good. A chronological split — earliest data to train, latest to test — simulates the real situation: you only ever have the past to predict the future.”

Q12. What would you do to improve this project? “A few things. I’d make the materials fully consistent — the write-up mentions a stationarity test, ARIMA confidence intervals, and VaR that I’d actually implement in the notebooks. I’d add proper mean-variance portfolio optimization to find efficient weights rather than using illustrative 60/30/10. I’d benchmark ARIMA against richer models. And I’d add walk-forward validation — retraining on a rolling window — since a single static split doesn’t reflect how forecasting works in practice.”

Q13. What’s the single most important finding? “That diversification across low-correlated asset classes genuinely reduces risk — my 60/30/10 portfolio cut volatility meaningfully and improved the Sharpe ratio versus all-stocks — but that it’s conditional: correlations rise in a crisis, so the benefit shrinks exactly when it’s most needed. The practical conclusion is that diversification works, but it has to be monitored, not assumed.”

16. How to Walk Through This Project Live

If asked to screen-share the notebooks, use this order:

1. State the structure first — “it’s a five-notebook pipeline, each phase chained through saved CSVs: data → returns → correlation → decomposition → forecast.”
2. Notebook 1 — show the yfinance download with retry logic and the data-quality check (zero missing values). Make the point: validate before you analyze.
3. Notebook 2 — walk the risk metrics. Explain Sharpe and the volatility-vs-drawdown distinction with the TLT example (moderate volatility, severe drawdown).
4. Notebook 3 — spend the most time here. It’s the strongest part. Show the correlation heatmap, the rolling correlations (equity clustering, the regime-dependent stock-bond relationship), and the 100%-SPY-vs-60/30/10 diversification comparison.
5. Notebook 4 — the decomposition; explain why multiplicative (exponential growth) and that SPY is trend-dominated.
6. Notebook 5 — the ARIMA baseline. Lead with the honesty: chronological split, a deliberately simple baseline, captured direction but not regime shifts.
7. Close on the diversification finding — it’s the most defensible, business-relevant conclusion. End on the insight, not the code.

Pacing tip: spend the most time on Notebooks 2 and 3 — risk metrics and diversification are the substantive finance content and the most defensible findings. Treat Notebook 5 as a baseline and be candid about its limits — that candor reads as maturity. And if the interviewer has read the portfolio page, be ready for §13: describe what the notebooks actually contain.

17. Glossary

• yfinance — an open-source Python library that pulls market data from Yahoo Finance; the project’s data source.
• OHLCV — Open, High, Low, Close, Volume — the standard daily price fields; this project uses Close.
• auto_adjust — a yfinance option that adjusts prices for stock splits and dividends.
• Ticker — a market symbol for an asset (SPY, AAPL, etc.).
• ETF — Exchange-Traded Fund — a basket of securities that trades like a stock (SPY, XLE, TLT, GLD, EFA are ETFs).
• Daily return — the day-over-day percentage price change; pandas pct_change().
• Cumulative return — the compounded growth of an initial $1; (1+r).cumprod().
• Total return — the whole-period percentage change, start to end.
• CAGR (annualized return) — Compound Annual Growth Rate; the constant yearly rate that yields the observed total return.
• Volatility — standard deviation of returns; the standard risk measure.
• Annualization (√252) — scaling a daily statistic to a yearly one using 252 trading days and the square-root-of-time rule.
• Sharpe ratio — risk-adjusted return: (return − risk-free rate) / volatility.
• Risk-free rate — the return on a “riskless” asset (a T-bill); this project uses 2%.
• Maximum drawdown — the worst peak-to-trough percentage decline over a period.
• Running maximum (cummax) — the highest value seen so far; used to compute drawdown.
• Correlation — how two assets co-move, from −1 to +1.
• Diversification — combining low-correlated assets to lower portfolio risk; a “free lunch.”
• Rolling correlation — correlation computed over a moving window (here, 252 days) to track how a relationship changes over time.
• Regime — a distinct market environment (bull, bear, crisis); relationships differ by regime.
• Stationarity — a series whose statistical properties are constant over time; returns are stationary, prices are not.
• Time-series decomposition — splitting a series into trend, seasonal, and residual components.
• Multiplicative vs. additive model — multiplicative for exponentially-growing series (markets); additive for linearly-growing ones.
• Moving average (50-day / 200-day) — a rolling-mean trend line.
• Golden cross / death cross — the 50-day MA crossing above / below the 200-day MA — a bullish / bearish regime signal.
• ARIMA(p,d,q) — AutoRegressive Integrated Moving Average; a classic univariate forecasting model. p = AR lags, d = differencing, q = MA lags.
• Differencing — subtracting consecutive values to make a series stationary (the “I” in ARIMA).
• Train/test split (chronological) — for time series, training on the earliest data and testing on the latest, never random.
• MAE / RMSE / MAPE — forecast error metrics: mean absolute error, root-mean-squared error, mean absolute percentage error.
• ADF (Augmented Dickey-Fuller) test — a statistical test for stationarity. (Described in the project write-up; not implemented in the committed notebooks — see §13.)
• VaR (Value at Risk) — an estimate of the worst expected loss at a given confidence level. (Listed as a project skill; not implemented in the committed notebooks — see §13.)

This study guide documents the project as built. The authoritative references are the five notebooks in notebooks/ (the actual code), the processed CSVs in data/processed/, and the portfolio page index.md. Where the portfolio page describes techniques not present in the committed notebooks, §13 documents the gap. When this guide and the notebooks disagree, the notebooks win.