See the market's next move before it's announced.

The present disclosure relates to data processing systems for telecommunications markets, and more particularly to computer-implemented systems that ingest heterogeneous signals about telecommunications operators, resolve them into a temporal knowledge graph, predict competitor actions using dependence-adjusted probabilistic fusion, prioritize resulting insights using a time-decay function, and evaluate counterfactual response strategies using subscriber-level simulation.

Telecommunications operators, including mobile network operators, mobile virtual network operators, cable and fiber broadband providers, and satellite direct-to-device providers, compete in markets where offers change frequently and switching friction is falling due to technologies such as embedded SIM (eSIM) provisioning.

Information relevant to competitor behavior is scattered across many unrelated sources: published tariffs and their fine print, regulatory and spectrum filings, radio equipment authorizations, mobile application releases, employment postings, domain and trademark registrations, public procurement notices, crowd-sourced network measurements, retail footprint changes, and social media. These sources differ in format, latency, and reliability.

Existing approaches include (a) periodic analyst reports, which have latency of weeks to months; (b) generic web-change monitoring tools, which produce high volumes of unranked alerts without telecom-specific semantics; and (c) network performance analytics products, which measure experience but do not model competitor intent. None of these combines the sources into a calibrated prediction of a specific competitor action within a time window.

Naively combining signals causes systematic overconfidence, because many signals share a common origin (for example, one press leak re-reported by dozens of outlets), which produces false alerts and alert fatigue. In addition, the value of a competitive insight decays quickly. An alert about a two-week promotion that reaches a decision-maker after three weeks has no value, yet conventional systems deliver alerts in first-in, first-out order or on fixed schedules.

There is therefore a need for a system that fuses heterogeneous telecom signals without double-counting shared origins, ranks insights by their decaying value, and quantifies response options before a decision is made.

In one aspect, a computer-implemented method ingests signals from a plurality of heterogeneous sources, resolves each signal to entities of a temporal telecommunications knowledge graph, and, for a candidate action class, computes a fused likelihood that an operator will perform the action within a predicted time window.

The fusion combines per-signal log-likelihood ratios weighted by learned source-reliability scores and discounted by a dependence factor derived from shared-origin clustering, so that correlated signals do not inflate the fused likelihood.

Each resulting insight is assigned a half-life based on the action class and predicted time window. A priority score decays exponentially with the insight's age, and a routing subsystem delivers the insight to a role-specific recipient before the priority score falls below a threshold.

For insights whose priority exceeds a simulation threshold, a market simulation engine runs a subscriber-level model calibrated on historical switching data to estimate the subscriber and revenue impact of a plurality of counterfactual response strategies. The ranked strategies are delivered with the insight. Running simulations only when triggered by events reduces compute consumption compared with continuous scenario modeling.

Further aspects include a tariff genome that decomposes offers into normalized attribute vectors to detect "stealth" changes, a grounded natural-language generator constrained to a provenance chain, a feedback calibrator that updates reliability scores from realized outcomes, and competition-law guardrails for multi-operator benchmarking.

FIG. 1 is a block diagram of a competitive-intelligence system 100 according to an embodiment.

FIG. 2 illustrates a tariff genome vector and detection of a stealth mutation.

FIG. 3 is a graph of a time-decaying priority score and a routing deadline.

System overview. Referring to FIG. 1, system 100 comprises a signal ingestion subsystem 110, an entity resolution engine 120, a temporal telecommunications knowledge graph 130, a tariff genome module 140, a fusion engine 150, an insight generator 160, a half-life scheduler 170, a market simulation engine 180, a routing and delivery subsystem 190, and a feedback calibrator 195. The components may run on one or more processors in a cloud or on-premises environment.

Ingestion 110. Connectors retrieve signals at source-specific cadences: tariff pages and terms documents (e.g., hourly), regulatory and equipment-authorization databases (daily), application store release artifacts including binary string tables and release notes (per release), employment postings (daily), domain and trademark registrations (daily), procurement notices (daily), crowd-sourced network measurements aggregated per hexagonal cell (hourly), and first-party data such as number port-out records (near real time). Each signal is stored with a timestamp, a source identifier, a content hash, and a canonical-origin fingerprint.

Entity resolution 120 and graph 130. Signals are linked to nodes representing operators, brands, offers, spectrum licenses and bands, sites and geographic cells, devices, enterprise accounts, executives and vendors. Edges carry validity intervals, so the graph can be queried "as of" any time. This lets the system reconstruct what was knowable at the moment a decision was made, which is used for back-testing.

Tariff genome 140. Each offer is converted into a genome vector g = (price, data allowance, hotspot allowance, standalone-5G access, bundled-perk value, roaming entitlement, contract lock-in, recurring fees, satellite messaging, device credit), with each gene normalized to [0, 1] in the direction of customer value. A stealth mutation is detected when at least one non-price gene changes by more than a threshold while the headline price gene is unchanged. The system raises a mutation event whose effective-price impact is computed from a hedonic regression across all offers in the market (FIG. 2).

Fusion 150. For an operator o and action class a (e.g., fixed-wireless launch, price reduction, market entry, partnership), the fusion engine starts from a prior log-odds L₀ learned from historical base rates and adds evidence from each enabled signal i:

where LR_i is the likelihood ratio of signal i given the action versus its absence, r_i ∈ [0, 1] is a learned reliability of the signal's source type, and d_i ∈ [0, 1) is a dependence discount. In one embodiment, signals are clustered by canonical-origin fingerprint and by source-type correlation learned from history. Within a cluster, the first signal receives d = 0 and each additional signal receives a positive discount (e.g., 0.5), so re-reported evidence cannot drive P toward certainty. The engine also outputs a predicted time window derived from the empirical lead-time distribution between each signal type and realized actions of class a, and a "decisive signal" identified as the signal whose removal most reduces P.

Insight generator 160. An insight record comprises the fused probability, the time window, an estimated impact I (e.g., revenue at risk), and a provenance chain listing every contributing signal with its identifier, source, timestamp, reliability and discount.

Half-life scheduler 170. Each insight receives a half-life h that depends on its action class and time window (e.g., hours for an outage-driven churn window, weeks for a spectrum consultation). Its priority is:

Simulation engine 180. When Priority exceeds a simulation threshold, the engine creates a market model with subscriber agents segmented by plan, tenure, device and eSIM capability, network-experience cell and price sensitivity, calibrated on historical port-in and port-out data. It runs the predicted action against a plurality of response strategies (e.g., hold, full match, partial match, non-price perk, targeted retention credits) and response lags. For each, it outputs incremental port-outs, customer-lifetime-value loss avoided, response cost and net impact. Strategies are ranked by net impact subject to user-configured constraints such as a maximum ARPU dilution.

Grounded narration. A language model generates a natural-language explanation of the insight and ranked strategies. The generator is constrained to statements supported by the provenance chain and simulation outputs. Each sentence carries at least one signal identifier, and sentences without support are removed before delivery.

Routing 190. Recipients are chosen by role (for example pricing, retail, network planning, enterprise sales, regulatory or executive) according to the action class, the impact and the time to t*. For insights above a critical threshold, the subsystem can automatically open a decision room in a collaboration tool, pre-filled with the evidence, the ranked strategies and proposed owners.

Feedback calibrator 195. When an action is realized or the window expires, the outcome is recorded. Reliability scores r_i, likelihood ratios, lead-time distributions and simulation elasticities are updated, for example by Bayesian updating or gradient-based calibration against a proper scoring rule such as the Brier score.

Guardrails. In a multi-operator benchmarking embodiment, contributed metrics are released only as aggregates across at least k contributors and after a minimum time lag. Forward-looking pricing information is excluded, so the system cannot serve as a channel for coordination between competitors. Personal data is excluded from social and first-party signals or aggregated before ingestion.

Technical effects. The dependence discount reduces false-positive alerts compared with independent-evidence fusion. Decay-bounded routing puts an upper bound on delivery latency for short-lived insights. Event-triggered simulation reduces processor time compared with continuous scenario modeling across all operators and action classes.

Alternatives. The fusion may instead use a learned model such as gradient-boosted trees or a graph neural network over graph 130, provided dependence between signals is accounted for. The decay function may be non-exponential. The simulation may be agent-based, system-dynamics or a learned surrogate model. The techniques also apply to adjacent regulated network industries such as energy retail and pay-TV.