Skip to content
Back to search results

Cross-Impact Analysis

Introduced by
Theodore Gordon & H. Hayward; Olaf Helmer
Year
1968
Type
Conditional-probability simulation of interacting future events
Family
Strategic management / strategic foresight
Font size

What does it do?

Cross-impact analysis is a forecasting technique that models how a set of possible future events influence one another, so that forecasts account for the fact that real events are interdependent rather than isolated.

Theodore Gordon and H. Hayward introduced the cross-impact matrix method in their 1968 Futures paper, motivated by the observation that judgmental forecasts such as Delphi estimate the likelihood of each event separately and ignore that the occurrence of one event can sharply raise or lower the odds of others. Olaf Helmer's 1977 work refined the approach, distinguishing the original correlational formulation from a causal cross-impact model and addressing the internal-consistency problems that plagued early matrices. The method specifies prior probabilities for events and conditional 'cross-impact' probabilities between them, then simulates the system to produce internally consistent joint outcomes and revised probabilities.

The intuition behind the method

Ask experts to estimate the chance of several future developments one at a time and you get a list of standalone probabilities - but the world does not work one event at a time. A breakthrough in battery technology changes the odds of electric-vehicle adoption, which changes the odds of new emissions regulation, which feeds back on the technology. Treating these as independent throws away the most important information: how they push on each other. Cross-impact analysis captures exactly that. Beyond each event's prior probability, experts estimate, for every pair, how the occurrence of one would change the probability of the other - the cross-impacts. The analyst then simulates the whole tangle many times: in each run, events occur or not according to their probabilities, and whenever one occurs it updates the others through the cross-impact links. After thousands of runs the frequencies settle into revised, mutually consistent probabilities and reveal which combinations of events tend to hang together. The payoff is a forecast that respects interdependence and a structured bridge between expert judgment and the internally consistent worlds that scenario planning needs.

History

Cross-impact analysis emerged from futures research in the late 1960s as a response to a known weakness of the Delphi method: that it forecasts events in isolation. Theodore Gordon and H. Hayward described the original cross-impact matrix method in 1968, reportedly growing out of work connected to a forecasting game, and proposed estimating conditional probabilities among events and simulating their interactions. Through the 1970s several variants appeared, and the approach drew criticism for internal inconsistency between elicited priors and conditionals. Olaf Helmer - a co-originator of Delphi - addressed these concerns in his 1977 Futures article, distinguishing correlational from causal cross-impact analysis and tightening the probabilistic foundations. Cross-impact methods became a standard tool in technology forecasting and corporate and governmental foresight, frequently paired with Delphi for input and with scenario planning to generate consistent futures, and later inspired variants such as cross-impact balance analysis.

When to use it

Use cross-impact analysis when a forecasting or foresight problem involves several future events that plausibly influence one another and treating them independently would be misleading, and when expert judgment is the main source of information. It is well suited to technology forecasting, policy and risk assessment, and especially to giving scenario planning a quantitative backbone by identifying which combinations of uncertainties are internally consistent. It works best with a small set of clearly defined events for which experts can credibly estimate pairwise conditional probabilities. The method is poorly suited to problems with many events (the matrix becomes unwieldy and the conditional estimates unreliable), to continuous variables rather than discrete events, and to situations lacking experts who can judge the interactions. Because it rests entirely on subjective priors and cross-impacts, it should be treated as a structured way to explore interdependence and consistency, not as an objective predictor.

Limitations and considerations

  • The number of pairwise conditional estimates grows quadratically with events, so the method scales badly beyond a small event set.
  • Eliciting reliable conditional probabilities is cognitively demanding and the estimates are often noisy or biased.
  • Separately elicited priors and cross-impacts are typically inconsistent and require non-trivial calibration to satisfy probability axioms.
  • Results depend entirely on subjective inputs and on the chosen propagation algorithm, limiting objectivity and reproducibility.

Source conditions do not replace independent scientific review or validation by an analysis engine.

Strengths

  • Captures the interdependence of future events that separate, single-event forecasts ignore.
  • Produces internally consistent revised probabilities and event combinations through simulation.
  • Bridges judgmental forecasting (such as Delphi) and scenario planning by quantifying which scenarios hang together.
  • Makes experts' assumptions about event interactions explicit, auditable, and testable via sensitivity analysis.

Applications

  • Adjusting technology-forecast probabilities to reflect how developments enable or block one another.
  • Generating internally consistent scenarios as a quantitative complement to intuitive-logics scenario planning.
  • Assessing how policy or regulatory events would cascade through interrelated future developments.
  • Risk analysis where the occurrence of one event changes the likelihood of correlated risks.
  • Refining Delphi forecasts by incorporating the cross-impacts among separately estimated events.

How it works

How it works

1

The analysis starts by defining a manageable set of distinct future events or developments relevant to the focal question and assigning each a prior probability of occurring within the time horizon, usually elicited from experts. Gordon and Hayward designed the method to extend judgmental forecasting, so these priors often come from a preceding Delphi exercise. The events must be clearly specified and reasonably independent in definition - overlapping or vaguely worded events corrupt the later conditional estimates - and the set is kept small because the number of pairwise interactions grows rapidly with the number of events.

Formula
E = \{E_1,\dots,E_n\},\quad P(E_i) = p_i \in [0,1] \ \text{(prior probabilities)}
2

Next, experts estimate the cross-impacts: for each ordered pair of events, the conditional probability of one event given that the other has occurred, recorded in a cross-impact matrix. This is the core of the method, capturing whether and how strongly the occurrence of event i would enhance or inhibit event j. Gordon and Hayward's central contribution was to make these interaction effects explicit and quantitative rather than leaving them implicit in separate forecasts. Each off-diagonal cell encodes both a direction (reinforcing or hindering) and a magnitude, turning a list of independent guesses into a structured model of how the future events are linked.

Formula
M_{ij} = P(E_j \mid E_i),\quad \text{matrix of conditional probabilities},\ i \ne j
3

Because the priors and the conditional estimates are elicited separately, they will generally not be mutually coherent under the laws of probability, and early cross-impact matrices were criticized for this. Helmer's refinement confronts the problem directly: the matrix must be checked so that the conditional probabilities and the marginals together satisfy the probability axioms, and inconsistent entries are adjusted - by the experts or by an algorithm - until the system is internally consistent. This calibration step is what separates a defensible cross-impact model from an arbitrary set of numbers, and it is where the causal formulation Helmer advocated improves on the original correlational version.

Formula
\sum \text{(implied marginals)} \ \text{must satisfy probability axioms};\ \text{adjust } M_{ij} \ \text{if violated}
4

With a consistent matrix in hand, the interactions are propagated by Monte Carlo simulation. In each of many synthetic runs, the events are visited in turn and 'occur' or not according to their current probabilities; whenever an event occurs, the cross-impact entries update the probabilities of the remaining events before they are sampled, so each run produces one coherent combination of occurred and non-occurred events. Gordon and Hayward's experiments showed that repeating this stochastic play-through thousands of times lets the web of conditional influences work itself out, generating a distribution over joint outcomes that no single-event forecast could provide.

Formula
\text{for } k=1..N:\ \text{sample events, update } p_j \leftarrow f(p_j, M_{ij}\mid E_i \text{ occurred})
5

Finally the simulation output is summarized. The revised probability of each event is the frequency with which it occurred across all runs, now reflecting the influence of the other events, and these adjusted marginals often differ markedly from the elicited priors. Just as importantly, the frequencies of particular event combinations identify the internally consistent scenarios the system favors. Helmer emphasized using these results both to sharpen probabilistic forecasts and to feed scenario construction, and analysts routinely run sensitivity tests - perturbing priors and cross-impacts - to see which assumptions most drive the conclusions.

Formula
\hat{p}_j = \frac{1}{N}\sum_{k=1}^{N} \mathbb{1}[E_j \text{ occurred in run } k],\ \ \text{rank scenarios by frequency}

Common pitfalls

  • Including too many or overlapping events, which makes conditional estimates unreliable and the matrix unmanageable.
  • Using the elicited priors and cross-impacts without checking and enforcing probabilistic consistency.
  • Confusing correlation with causation in the cross-impact estimates, the very issue Helmer's causal formulation sought to fix.
  • Reporting the revised probabilities as objective forecasts rather than as conditional outputs of subjective expert judgment.

Frequently asked questions

What problem with Delphi does cross-impact analysis solve?

Delphi and similar judgmental methods estimate the probability of each event on its own, implicitly treating events as independent. Gordon and Hayward observed that future events actually enable or inhibit one another, so independent estimates can be jointly incoherent and miss important dynamics. Cross-impact analysis adds a matrix of conditional probabilities describing how each event affects the others and simulates the interactions, yielding revised probabilities and consistent event combinations. It is therefore often run on the output of a Delphi study to repair exactly this blind spot.

How does the simulation turn a matrix into forecasts?

The cross-impact matrix is propagated by Monte Carlo simulation. In each run, events are sampled as occurring or not according to their probabilities, and whenever an event occurs the matrix updates the probabilities of the remaining events before they are sampled, producing one internally consistent combination. Repeating this thousands of times, as in Gordon and Hayward's experiments, gives the frequency with which each event occurs - the revised probability - and the frequency of each event combination, which identifies the scenarios the system favors.

Why do early cross-impact matrices need a consistency check?

Because the prior probabilities and the pairwise conditional probabilities are elicited separately from experts, they usually do not satisfy the laws of probability simultaneously - implied marginals may contradict the stated priors. Helmer's 1977 work made this a central concern, introducing a causal formulation and calibration so the matrix obeys the probability axioms before simulation. Skipping this step produces results that look quantitative but rest on an incoherent probability model, so consistency checking is essential to a credible cross-impact analysis.

Related methods

Sources

  1. Gordon, T. J., & Hayward, H. (1968). Initial experiments with the cross impact matrix method of forecasting. Futures, 1(2), 100-116.

    Open DOI reference
    Copied.
  2. Helmer, O. (1977). Problems in futures research: Delphi and causal cross-impact analysis. Futures, 9(1), 17-31.

    Open DOI reference
    Copied.