1.3 Time Series Analysis, Monte Carlo Simulation & Machine Learning

1.3 Time Series Analysis, Monte Carlo Simulation & Machine Learning

1. Decomposing a Time Series

A financial time series is conventionally decomposed into four components:

The distinction between seasonal and cyclical matters on the exam: seasonality has a fixed period (monthly, quarterly, annual) and can be removed by seasonal adjustment; cyclical variation has variable duration and amplitude and cannot.

Smoothing techniques suppress noise to reveal the underlying signal:

• A simple moving average of order $k$ averages the last $k$ observations, giving each equal weight.
• Exponential smoothing weights recent observations more heavily, with weights declining geometrically:

$S_t = \alpha Y_t + (1 - \alpha) S_{t-1}$

A higher $\alpha$ makes the smoothed series more responsive to recent data; a lower $\alpha$ produces a more stable but more lagged signal. All smoothing methods introduce lag — the price paid for noise reduction.

2. Mean Reversion and Multi-Period Risk

A series is mean-reverting when deviations from a long-run average tend to be followed by moves back toward it, which shows up statistically as negative serial correlation at longer horizons. This has a direct and heavily tested portfolio consequence.

Under the assumption that returns are independent and identically distributed (i.i.d.), multi-period volatility scales with the square root of time:

$\sigma_T = \sigma_{1} \times \sqrt{T}$

If returns instead mean-revert, long-horizon variance grows more slowly than $T$, so:

• Long-horizon risk is lower than square-root-of-time scaling implies.
• The case for higher equity allocations at long horizons is strengthened.
• Conversely, positive serial correlation (momentum, or smoothed appraisal-based valuations) makes long-horizon risk higher than the naive scaling suggests.

The variance ratio test formalizes this: it compares the variance of $T$-period returns to $T$ times the variance of one-period returns. A ratio below 1.0 indicates mean reversion; above 1.0 indicates trending behavior.

Exam trap: mean reversion in valuation ratios (such as CAPE) is far better documented than mean reversion in realized returns. Candidates should not assume that because an asset has fallen, its expected return has mechanically risen by an equal amount.

3. Monte Carlo Simulation

Monte Carlo simulation replaces a single deterministic projection with thousands of randomized paths. The procedure:

1. Specify the model inputs — expected return, volatility, and the correlation matrix across asset classes, plus the cash-flow schedule (contributions, withdrawals, spending policy).
2. Choose a distributional assumption — normal, lognormal, Student's t (for fatter tails), or bootstrapped historical blocks.
3. Generate a path: draw a random return for each period and compound forward, applying cash flows in the correct sequence.
4. Repeat 1,000 to 100,000 times.
5. Rank the terminal outcomes and read off the distribution.

The output is a distribution, and its central planning statistic is a probability of success — the share of paths in which the portfolio funds all required cash flows without depletion.

Monte Carlo's decisive advantage over a straight-line projection is that it captures sequence-of-returns risk: for a portfolio in distribution, the order of returns changes the outcome even when the average return is identical, because withdrawals during early drawdowns permanently remove capital that cannot participate in a later recovery.

Limitations the exam expects candidates to state:

• Garbage in, garbage out. Simulation refines arithmetic, not forecasting. Inputs drawn from a favorable historical window propagate directly into the answer.
• Distributional misspecification. Normal assumptions understate tail frequency; real returns are negatively skewed with excess kurtosis.
• Static correlations. Most implementations hold the correlation matrix fixed, whereas correlations rise toward 1.0 in crises — precisely when diversification is needed.
• False precision. A "87.3% probability of success" implies a confidence the model does not possess. Practitioners should treat bands (e.g. 80–90%) as decision-relevant, not the third digit.

4. Machine Learning in Investment Analysis

The current CIMA blueprint explicitly names machine learning among the computational algorithms whose outputs candidates must interpret.

Practical uses in wealth and investment consulting include natural-language processing of filings, transcripts and news for sentiment; alternative-data analysis such as satellite imagery or card-transaction data; document review in operational due diligence; and anomaly detection in trade surveillance and compliance.

Why financial data is unusually hostile to machine learning:

1. Low signal-to-noise ratio. Asset returns are close to unpredictable; most variation is noise that a flexible model will happily fit.
2. Overfitting and backtest overfitting. A model with enough parameters can fit any historical sample perfectly and fail completely out of sample. Guarding against this requires out-of-sample testing and walk-forward validation, in which the model is repeatedly trained on data strictly preceding the period it is tested on.
3. Non-stationarity. Market relationships change with regime, policy, and structure, so a relationship learned in one decade may not survive into the next.
4. Short effective samples. Fifty years of monthly data is only 600 observations — trivially small for models with hundreds of parameters.
5. Interpretability and fiduciary duty. A model whose recommendations cannot be explained is difficult to reconcile with the duty to act in a client's best interest and to disclose the basis for advice. "Black-box" outputs still require documented human oversight.

The consulting standard is therefore that machine-learning output is evidence to be weighed, not a decision rule to be followed, and that any model driving client portfolios must be documented, validated out of sample, and monitored for degradation.