Dynamic Mortality Forecasting via Mixed-Frequency State-Space Models

summary

The gist

Dynamic Mortality Forecasting via Mixed-Frequency State-Space Models proposes a mixed-frequency state-space (MF–SS) extension of the Lee–Carter framework that jointly uses annual mortality rates…

In short

This research introduced a Mixed-Frequency State-Space (MF–SS) model to forecast mortality by jointly using annual rates and monthly death counts. Unlike traditional methods, this model allows for real-time updates of forecasts as new monthly data arrives without re-estimating the entire model. It proves that pooling these frequencies provides more accurate point forecasts and better uncertainty estimates than using separate annual or monthly models.

Key concepts

Mixed-Frequency State-Space (MF–SS) Model
This is a statistical framework that combines two different time scales—annual and monthly—into one model. It links them through a shared hidden factor, allowing the system to use both types of data simultaneously. This structure enables dynamic updates when new monthly data becomes available.
Temporal Forecast Reconciliation
This technique takes forecasts from two separate models (one annual, one monthly) and combines them into a single forecast. It forces the combined result to obey real-world rules, such as ensuring the sum of monthly deaths equals the annual total, improving accuracy when combining different time resolutions.
Latent Monthly Mortality Factor
This is an unobserved hidden variable within the model that captures short-term dynamics like seasonality and immediate trends in mortality. The model uses a Seasonal ARIMA process to describe this factor, allowing it to capture both long-term changes and short-run fluctuations.
Intra-year Updates
This is the ability of the MF–SS model to instantly refine its predictions for the current year as new monthly observations are recorded. Because only a small part of the model needs updating, this feature supports dynamic learning loops, allowing AI systems to incorporate fresh data immediately.

Terminology used across episodes

This episode discusses

The paper

Dynamic Mortality Forecasting via Mixed-Frequency State-Space Models · Read on arXiv

Runze Li, Rui Zhou, David Pitt

Department of Economics, University of Melbourne

High-frequency death counts are now widely available and contain timely information about intra-year mortality dynamics, but most stochastic mortality models are still estimated on annual data and therefore update only when annual totals are released. We propose a mixed-frequency state-space (MF--SS) extension of the Lee--Carter framework that jointly uses annual mortality rates and monthly death counts. The two series are linked through a shared latent monthly mortality factor, with the annual period factor defined as the intra-year average of the monthly factors. The latent monthly factor follows a seasonal ARIMA process, and parameters are estimated by maximum likelihood using an EM algorithm with Kalman filtering and smoothing. This setup enables real-time intra-year updates of the latent state and forecasts as new monthly observations arrive without re-estimating model parameters. Using U.S. data for ages 20--90 over 1999--2019, we evaluate intra-year annual nowcasts and one- to five-year-ahead forecasts. The MF--SS model produces both a direct annual forecast and an annual forecast implied by aggregating monthly projections. In our application, the aggregated monthly forecast is typically more accurate. Incorporating monthly information substantially improves intra-year annual nowcasts, especially after the first few months of the year. As a benchmark, we also fit separate annual and monthly Lee--Carter models and combine their forecasts using temporal reconciliation. Reconciliation improves these independent forecasts but adds little to MF--SS forecasts, consistent with MF--SS pooling information across frequencies during estimation. The MF--SS aggregated monthly forecasts generally outperform both unreconciled and temporally reconciled Lee--Carter forecasts and produce more cautious predictive intervals than the reconciled Lee--Carter approach.

Transcript

Introduction to the show: ident: Economics. The day in economics: theory, mechanisms and the evidence beneath them.

Ruth: We're looking at the paper: "Dynamic Mortality Forecasting via Mixed-Frequency State-Space Models".

Warren: Here's the gist of the paper. Dynamic Mortality Forecasting via Mixed-Frequency State-Space Models proposes a mixed-frequency state-space (MF–SS) extension of the Lee–Carter framework that jointly uses annual mortality rates and monthly death counts to enable real-time intra-year updates and forecasts without re-estimating model parameters Model Architecture The core contribution is a mixed-frequency state-space (MF–SS) mortality model that extends the Lee–Carter structure to jointly use annual mortality rates and monthly death counts, linking annual and monthly dynamics through a shared latent monthly mortality factor.

Ruth: This framework includes two observation equations—one for annual log mortality rates and one for a scaled monthly log series constructed from monthly deaths—linked through a shared latent monthly mortality factor.

Warren: The annual period factor is defined as the intra-year average of the monthly factors, providing a structural cross-frequency link between annual and monthly dynamics.

Ruth: The latent monthly factor follows a seasonal ARIMA (SARIMA) process, capturing long-run drift, short-run persistence, and seasonal dynamics.

Title and authors: Ruth: So this paper introduces something called Dynamic Mortality Forecasting via Mixed-Frequency State-Space Models. It’s trying to link two different types of data to get better forecasts than what we usually see.

Warren: I saw the title, and it sounds complicated, but basically, they’re looking at how to use both annual mortality rates and monthly death counts together. That sounds like a real challenge for getting timely information into the models.

Iris: It seems like the main idea is to use a mixed-frequency state-space model that connects these two series through one shared latent monthly mortality factor. This factor is what’s doing all the linking between the annual and monthly dynamics.

Ruth: Exactly, and they define an annual period factor as being the average of those monthly factors during the year. That way they create a structural link between what we see annually and what we see month-to-month.

Warren: The latent monthly factor itself follows a seasonal ARIMA process, which lets them capture both the long-run trend and any short-term bumps within those months. It’s a way to model that underlying noise in the monthly data.

Iris: And they estimate all these parameters using maximum likelihood through an EM algorithm that uses Kalman filtering and smoothing. This setup is designed specifically so they can get real-time updates as new monthly observations come in without having to re-estimate all those big model parameters every time.

Ruth: That’s the key selling point there, isn't it? The ability to update the latent state dynamically intra-year while keeping the main model parameters stable until a full additional calendar year of data has passed.

Warren: Right, and they use U.S. data for ages twenty to ninety over that period from nineteen ninety nine up to two thousand nineteen in their tests. That gives them a solid dataset to work with when they evaluate how accurate these forecasts are for both intra-year annual nowcasts and those one-to-five-year ahead predictions.

Iris: When they look at the results, they found that incorporating this monthly information materially improves those intra-year annual nowcasts, especially after the first few months of the year when things are still pretty sparse.

Title and authors: Ruth: That makes sense from an economist's view; having more frequent data lets you capture immediate shifts in mortality patterns rather than just waiting for the annual release. But what about the numbers?

Warren: Well, they found that their aggregated monthly forecast, which comes from this mixed-frequency model, is typically more accurate than just a direct annual forecast or even those forecasts made by combining separate Lee-Carter models.

Iris: They noted that this aggregated monthly forecast is usually more accurate than the unreconciled and temporally reconciled Lee-Carter forecasts they built separately, which suggests that pooling the information across frequencies during estimation really helps them.

Ruth: That points toward why they chose this mixed-frequency structure; it’s about combining different resolutions of data in a way that maximizes predictive power for point estimates.

Warren: And they found that even when comparing the state space model against other methods, the MF-SS model still beats those temporally reconciled Lee-Carter forecasts, which is interesting because reconciliation usually does a good job of forcing the annual and monthly forecasts to match up on the aggregation constraint.

Iris: That comparison suggests that while reconciliation helps independently estimated models get better, pooling information through the shared latent factor in this paper’s MF-SS framework is what really gives it an edge when you care about both point accuracy and a more cautious assessment of uncertainty.

Ruth: So, if I were looking at this for market modeling, it means we can use high-frequency data immediately to adjust our expectations for annual trends without waiting for the full year's data to come in.

Warren: And from a trader's side, that real-time update capability is useful because it reduces the lag between a mortality shock happening now and our ability to incorporate that into our forward-looking risk assessment.

Iris: I just want to check on the identification again—what exactly is doing the work here? The paper shows that for both annual and monthly factors, they are nearly indistinguishable in terms of their baseline age-specific mortality level.

Ruth: That’s a good observation; it means the fundamental age structure component isn't drastically different between the two frequencies, which makes sense since they are built on the same underlying mortality process.

Title and authors: Warren: They did find that the period factors do show pronounced short-run oscillations consistent with seasonality, while those annual factors are smoother and capture that long-run decline in mortality over the sample. That’s a clear distinction in how each frequency behaves.

Iris: And when we look at uncertainty, they found that the prediction intervals produced by this state space model are much wider than those from the reconciled Lee-Carter approach for both ages.

Ruth: That makes sense because they noted that the estimated innovation variance of the monthly period factor under this state space specification is significantly larger than in their other model. So, if you need a very cautious view, you get wider intervals but potentially more reliable point forecasts.

Warren: And that wider interval is driven by this larger state innovation variance accumulating over the forecast horizon, which means those uncertainty bands stretch out as you look further into the future.

Iris: They also pointed out a limitation: the paper states that their method does not produce corresponding monthly forecasts directly; it targets annual outcomes only.

Ruth: That’s a fair point; they aren't giving you a direct monthly projection, just an implied annual one derived from those high-frequency dynamics.

Warren: But they did suggest extensions for the future, like modeling cause-specific mortality jointly across different causes or even multi-country modeling where data coverage and reporting frequency differ.

Iris: They also suggested improving reconciliation methods by incorporating cross-frequency error covariances to better reflect dependence between the annual and monthly resolutions.

Ruth: So, to wrap up on the implications of Dynamic Mortality Forecasting via Mixed-Frequency State-Space Models, this paper shows that when you want both point accuracy and a more cautious assessment of uncertainty, this mixed-frequency approach is a way to pool information across frequencies during estimation.

Warren: It moves beyond just fitting separate models and instead builds one joint structure that handles the misalignment between annual and monthly data naturally by treating unobserved observations as missing data.

Iris: It’s a practical tool for getting timely updates without constantly re-estimating all the model parameters, which is crucial when you need to react quickly to new information like those monthly death counts.

Ruth: And that's what I think, for now. We're going to move on from this paper and see what other data challenges they tackle next.

The paper's summary: Ruth: So, this paper proposes a method to forecast mortality by mixing annual rates and monthly death counts together using a special state-space model that lets us update things in real time without constantly re-estimating the main parameters.

Warren: It's essentially taking the Lee–Carter idea and adding a monthly layer to it, linking those two frequencies through one shared hidden factor that captures the underlying mortality trend.

Iris: The core contribution is this mixed-frequency state-space model structure where we have two observation equations—one for the annual log mortality rates and one for a scaled monthly series made from death counts—which are both linked by that single latent monthly factor.

Ruth: And they define an annual period factor as being the average of those monthly factors over a year, which creates this structural link between the yearly view and the month-to-month view.

Warren: That latent monthly factor itself follows a seasonal ARIMA process, which lets them capture both long-term drift and short-run persistence in that monthly data.

Iris: The estimation uses an EM algorithm with Kalman filtering and smoothing, which is key because it lets us get real time updates of the latent state as new monthly observations arrive without having to re-estimate the main model parameters every single time.

Ruth: So, the big win here is this dynamic update capability; they can incorporate new data immediately into the forecast process while keeping the overall model structure stable until a full additional year of data has passed.

Warren: The benchmark they compare this against involves fitting two separate Lee–Carter models—one for annual and one for monthly—and then combining their forecasts using temporal reconciliation, which forces them to obey the aggregation constraint that annual deaths equal the sum of monthly deaths.

Iris: They found that temporal reconciliation does improve those separately estimated forecasts, but it doesn't add much to the performance of this mixed-frequency model because they're already pooling information across frequencies during estimation.

Ruth: That suggests that when you care about both getting a good point estimate and also having a more cautious view of how wrong your prediction might be, this approach is preferred.

Warren: The empirical results on U.S. data show that incorporating this monthly information materially improves the intra-year annual nowcasts, especially after the first few months of the year when those monthly observations are still sparse.

Iris: They also noted that their aggregated monthly forecast is typically more accurate than both the unreconciled and temporally reconciled Lee–Carter forecasts, even producing more cautious prediction intervals than the reconciled approach.

Ruth: That difference in uncertainty is important; they found that the state space model produces much wider prediction intervals because the estimated innovation variance of that monthly period factor under this specific model is significantly larger.

Warren: So, to summarize, this mixed-frequency state-space model offers real time updates and a way to pool information across resolutions for better point accuracy and more reliable uncertainty assessments.

Iris: And looking ahead, the authors suggest extensions like modeling cause-specific mortality jointly across different causes or even multi country modeling where data coverage varies between places.

Ruth: They also think we should improve reconciliation methods by adding cross-frequency error covariances to better capture the dependence between the annual and monthly time resolutions.

Warren: It’s a practical framework for AI systems because it handles data misalignment naturally by treating missing observations as just that—missing—and continues updating the latent states with whatever information is available.

Iris: This means for anyone building a system that needs to react quickly to new data, this structure provides a robust way to integrate high-frequency inputs into long-term trend modeling without needing constant re-training of the entire model.

Ruth: We’ll be talking next about how this dynamic updating plays out in practice and what it means for making better risk assessments with mortality data.

The paper's improvements: Ruth: So, we’re looking at what the authors suggest for next steps after showing off this mixed-frequency state space model for mortality forecasting.

Warren: They aren't just stopping at that; they actually propose ways to make the whole framework more robust.

Iris: They suggest extending the model to incorporate cause-specific mortality jointly across different causes, which would let us account for shared period dynamics and also specific deviations related to certain diseases.

Ruth: That’s interesting because it moves beyond just the general trend; it lets you look at how different health shocks affect mortality separately while still benefiting from that common underlying factor.

Warren: From a quantitative standpoint, that would mean we could build models for specific disease outbreaks and see how they interact with the overall population trends captured by the shared latent factor.

Iris: They also point toward multi country modeling where the data coverage and even the reporting frequency might differ between different nations.

Ruth: That’s a practical application; it acknowledges that mortality trends aren't uniform across geography, which is a big caveat in any global model.

Warren: And they suggest improving the reconciliation methods themselves by adding cross-frequency error covariances to better reflect the dependence between those annual and monthly resolutions when combining forecasts.

Iris: That’s an econometric improvement; it means when we combine different time scales, we should be explicitly modeling how errors in one frequency influence errors in the other.

Ruth: It sounds like they’re trying to build a more sophisticated way to handle the inherent noise and relationships between different data resolutions rather than just forcing them to match up at the end.

Warren: So, the implication for us as traders is that if we use this framework, we might get more realistic risk assessments because it accounts for how different types of shocks propagate through both time scales.

Iris: That makes sense; it’s about getting better calibration when you’re trying to predict something that involves multiple layers of time and cause.

Ruth: The underlying assumption here is that these extensions will help us model complex health shocks or specific disease trends while still benefiting from the general mortality trend captured by that common latent factor.

Warren: So, we’re moving from just a point forecast to a framework that can handle more complex, multi-faceted risk scenarios.

Iris: That's where the next area of research lies—making those extensions actually work well in practice when you have messy, real world data.

Ruth: And the challenge is always making sure that adding these new layers doesn't just add complexity without actually improving the predictive power significantly.

Conclusion: Tom: So, we're wrapping up our look at "Dynamic Mortality Forecasting via Mixed-Frequency State-Space Models," which essentially shows how you can combine annual and monthly death count data using a state space model for real time updates without constantly re-estimating everything.

Ruth: The big implication here is that for market incentives, it means we can get more immediate feedback on mortality trends as they happen, not just waiting for the next year's report.

Warren: Right, and from a trading side, it suggests a way to adjust your risk view intra-year when new monthly data comes in instead of having to wait for the full annual cycle.

Iris: I think the key identification here is that while both annual and monthly factors look different in their short term, they share a stable baseline mortality level that's robust across both frequencies.

Ruth: So, even if you change your assumptions about people's behavior or incentives, the fundamental age structure component of the model seems pretty solid.

Warren: That’s good for running large simulations; it means you don't have to constantly rebuild your entire population structure when you adjust a parameter related to long term shifts.

Iris: And while they show this pooling works well for point accuracy, they also flag that the prediction intervals get much wider because the model has a larger state innovation variance than other methods.

Ruth: So, we get more precise numbers for the point estimate, but we have to accept a wider band around it when making decisions.

Warren: That wide interval is what you’re paying for if you want that real time update capability; it tells you the range of uncertainty in that immediate period.

Iris: Ultimately, this paper shows how pooling information across frequencies during estimation can lead to better point accuracy and a more cautious assessment of uncertainty when predicting mortality.

Ruth: So, to wrap up on "Dynamic Mortality Forecasting via Mixed-Frequency State-Space Models," it’s a sophisticated way to integrate different data resolutions for dynamic risk modeling.

Warren: It’s useful stuff for anyone building an AI system that needs to react quickly to new information streams without needing a full re-calibration every single day.

Iris: It really highlights how the structure of the state space model can handle misalignment between annual and monthly data naturally by treating unobserved observations as missing.

Ruth: Next time, we’ll take a look at some of those extensions they proposed, like modeling cause specific mortality jointly across different causes.

More episodes

← Home