Time-Series Forecasting: Separating Predictive Skill From Hindsight

A financial time series carries information in its ordering. The same observations shuffled into a different sequence may have the same mean and variance but a completely different forecasting problem. Good forecasting therefore begins with the time structure of the data, not with a search for the most elaborate model.

This reference joins decomposition, autoregression and forecast evaluation into one workflow. It focuses on conditional-mean forecasts and their validation; volatility dynamics require their own modelling decisions. No historical teaching forecast is presented as a current investment call.

Define the clock and the target

Observations may arrive at regular intervals or irregular event times. A daily close, a monthly economic release and a transaction-level price stream have different information clocks. Before estimating a model, specify the observation time, release time and revision policy for every variable.

A target should describe what will be forecast and when that prediction becomes actionable. “Next day’s return from today’s close” is different from “tomorrow’s closing level using information available tomorrow morning.” If a variable is published after the forecast decision, its period label does not make it available earlier.

Price levels, simple returns and log returns also answer different questions. For positive prices $P_t$, log return is $r_t=\log(P_t/P_{t-1})$. It is additive across consecutive periods, but its mean is not the arithmetic mean return. Differencing a trending series can help stabilise its mean; it does not automatically stabilise variance or remove structural breaks.

Decomposition is a diagnostic rather than a causal theory

An additive representation writes an observed series as trend, seasonal, cyclical and irregular components. A multiplicative representation is useful when component magnitudes scale with the series level; logarithms can turn a positive multiplicative representation into an additive one.

These decompositions reveal useful features. Seasonal effects repeat on a known calendar, while a broader cycle need not have a fixed duration. Outliers may indicate data errors, genuine shocks or a change in regime. Smoothing choices determine what is labelled trend versus cycle, so the resulting components are not uniquely revealed economic truths.

For prediction, decompositions must be estimated within each historical training window. A smooth trend estimated using the whole dataset can use future observations to revise earlier trend values. Feeding that trend into a backtest creates look-ahead even if the forecasting regression itself is correctly dated.

Stationarity supports stable parameter interpretation

Weak stationarity means a constant mean, finite constant variance and autocovariance depending on lag rather than calendar date. It does not require independent observations. An autoregression can be stationary while exhibiting substantial dependence.

The autocorrelation function measures association between a series and its lagged values. Partial autocorrelation isolates the linear association at a lag after controlling for intermediate lags. In a theoretical autoregression of order $p$, the population partial autocorrelation cuts off after lag $p$; sample estimates fluctuate and rarely produce a perfect cutoff.

Persistent levels can make unrelated variables appear statistically connected. A visually high regression fit between trending series is therefore weak evidence of a useful predictive relationship. Differencing, detrending and, where appropriate, a properly specified long-run relationship address different forms of nonstationarity and should not be applied interchangeably.

An autoregression turns persistence into a forecast

An AR(1) model is

$$X_t=c+\phi X_{t-1}+\varepsilon_t,$$

where $c$ is an intercept, $\phi$ is persistence and $\varepsilon_t$ is an innovation with conditional mean zero. Under standard finite-variance assumptions, $|\phi|<1$ gives a stationary mean $\mu=c/(1-\phi)$.

The $h$-step conditional mean is

$$\widehat X_{t+h\mid t}=\mu+\phi^h(X_t-\mu).$$

An illustrative series with $\mu=2$, $\phi=0.8$ and current value 3 has next-step forecast 2.8 and two-step forecast 2.64. These numbers demonstrate conditional mean reversion. Realised observations can move farther from the mean because innovations continue to arrive.

If innovation variance is $\sigma_\varepsilon^2$, forecast-error variance for known parameters is

$$\operatorname{Var}(X_{t+h}-\widehat X_{t+h\mid t}) =\sigma_\varepsilon^2\sum_{j=0}^{h-1}\phi^{2j}.$$

In words: Forecast uncertainty accumulates the uncertainty of future shocks, with each shock weighted by how long its effect persists.

Parameter estimation adds uncertainty beyond this formula. Reporting only the model-conditional interval can materially understate uncertainty in a short sample or near a unit root.

For AR models of higher order, stationarity depends on the roots of the autoregressive lag polynomial, not on each coefficient being less than one individually. Adding more lags also increases the number of estimated parameters and can reduce forecast reliability when the sample is limited.

Moving-average shocks and differencing

A moving-average model of order $q$ uses current and past innovations. “Moving average” here does not mean an ordinary rolling average of observed prices. The innovations are model errors whose estimation and invertibility require their own treatment.

ARMA combines autoregressive dependence and moving-average shocks for an appropriate stationary series. ARIMA adds differencing. Its orders $(p,d,q)$ describe autoregressive lags, differencing order and moving-average lags.

Differencing can remove a stochastic trend, but unnecessary differencing can amplify noise and introduce an avoidable moving-average structure. Model selection should therefore combine data diagnostics, economic interpretation and out-of-sample performance instead of treating differencing as a universal cleaning step.

Several series increase both information and risk

A vector autoregression models a vector $\mathbf X_t$ using lagged vectors:

$$\mathbf X_t=\mathbf c+A_1\mathbf X_{t-1}+\cdots+A_p\mathbf X_{t-p}+\boldsymbol\varepsilon_t.$$

In words: Predict each series using a fixed term, past values of the series in the system, and a new unpredictable change.

The matrices $A_j$ describe lagged predictive relationships. The model may use interactions among rates, inflation and activity, but a predictive coefficient is not automatically a structural causal effect.

With $k$ variables and $p$ lags, each equation already has $kp$ lag coefficients before other controls. A larger system can exhaust limited observations quickly. It also requires genuinely available future inputs when used in a conditional forecasting exercise. Supplying realised future explanatory variables converts a forecast into an ex-post explanation.

Evaluate on the information available at the time

Randomly shuffling a time series into training and test samples can expose the model to future regimes and overlapping information. A more defensible evaluation uses rolling or expanding windows: fit on earlier observations, forecast the next period, record the error and advance the origin.

An expanding window retains all earlier data and can improve precision when relationships are stable. A rolling window discards older observations and can adapt more quickly to change. Neither dominates in every setting. The window length is itself a choice that should be selected without repeatedly optimising against the final test sample.

Multi-period targets can overlap. For example, adjacent forecasts of a one-month return share much of the same realised period. Their errors are dependent, so conventional independent-error confidence statements can be misleading. Data gaps, publication delays and embargoes between training and test observations may also be needed when labels overlap.

Choose the loss function before choosing the winner

For actual outcomes $y_i$, forecasts $\hat y_i$ and $n$ observations,

$$\operatorname{MAE}=\frac1n\sum_i|y_i-\hat y_i|, \qquad \operatorname{RMSE}=\sqrt{\frac1n\sum_i(y_i-\hat y_i)^2}.$$

In words: Average absolute misses, or average squared misses and take the square root to give large misses greater weight.

MAE reports typical absolute error in the original units. RMSE puts greater weight on large misses. Percentage errors can become unstable or undefined when actual values are close to zero, making them poor defaults for near-zero returns or spreads.

Evidence needed to evaluate a forecast

A lower statistical error does not automatically imply a profitable trading rule. A forecast can be directionally helpful yet too small to cover costs, or accurate in quiet periods while failing when positions matter most.

Compare forecasts on identical observations

A model comparison is valid only when the models face the same target dates and information constraints. If one model is evaluated after discarding difficult missing-data periods while another is evaluated throughout, their aggregate scores answer different questions.

An illustrative three-period comparison makes the metric choice visible.

Suppose Model A has errors of

Model A error magnitudes over the three listed periods

units,

while Model B has errors of

Model B error magnitudes over the three listed periods

Both have MAE of 2. Model A has RMSE of approximately 2.45; Model B has RMSE of 2. If large misses are disproportionately costly, the second error pattern is preferable under squared-error loss despite equal average absolute error.

Neither pattern says whether forecasts are systematically too high or too low because only error magnitudes were specified. Bias requires signed errors. Directional accuracy requires the correct comparison with a directional target. A single score cannot answer all three questions.

Prediction intervals need both calibration and sharpness

A 90% prediction interval should contain approximately 90% of realised outcomes over a suitable evaluation set when the model is well calibrated. Very wide intervals can achieve high coverage while providing little decision value. Coverage and width should therefore be reviewed together.

Conditional coverage matters as well. An interval that works in quiet periods and fails during stress may have acceptable average coverage while being least reliable when most needed. Evaluation by volatility regime, horizon and forecast magnitude can reveal that pattern.

Parameter uncertainty, innovation uncertainty and future regime uncertainty are different sources of error. A textbook autoregressive interval often includes only the innovation component conditional on estimated parameters. Resampling or more explicit parameter modelling can address part of the gap, but no mechanical adjustment guarantees protection against an unmodelled structural break.

Residual checks identify what the model has missed

After fitting a conditional-mean model, inspect residual autocorrelation. Persistent residual dependence suggests that the selected lags or variables have not captured all predictable linear structure. It can also reflect timing errors or an inappropriate transformation.

Inspect squared residuals separately. Little residual autocorrelation does not imply constant conditional variance. A model may be adequate for the mean and inadequate for uncertainty intervals. That is a reason to separate the mean specification from volatility modelling rather than demand that one autoregression solve both tasks.

Large residuals should be checked against original observations. Correcting an input error is appropriate; deleting a genuine crisis observation merely because it hurts fit changes the population the model claims to describe. The choice should be recorded rather than hidden inside generic outlier removal.

Updating changes the model being evaluated

A fixed-parameter forecast and a regularly re-estimated forecast are different procedures. If the intended operation re-estimates monthly, the historical evaluation should do the same. A backtest fitted once using the final sample does not represent that deployment.

The same applies to model selection. Choosing the best lag order separately at each historical origin can be valid if the choice uses only preceding data. Choosing the final best lag order after seeing the whole test period gives an optimistic assessment. The evaluation unit is the entire forecasting procedure, including its updating and selection rules.

Preserve the forecast record

Store the forecast origin, training window, data version, model parameters, benchmark and realised outcome. Separate the model-selection sample from the final evaluation period. Inspect residual autocorrelation and changes in error distribution rather than relying on a single aggregate score.

The objective is a repeatable conditional forecast with a clear uncertainty statement. A modest model that survives honest timing and benchmark checks is more useful than an elaborate retrospective fit whose apparent accuracy depends on information unavailable at the decision date.