Because all we can do is project from the patterns we already have seen, when unprecedented patterns emerge (i.e., patterns we have never seen before), our forecasting methods fail. This is an unavoidable reality. Short of clairvoyance, it couldn’t be any other way. It is impossible for the projected trends and confidence limits our tools produce to make reasoned allowances for types of changes we have never before experienced or thought about. In the context of our experience they are unexpected outliers.

Some series, especially tourism related, have reacted to covid-19 by producing patterns we have never seen before. For those situations the forecasts from our simple Year-13 tools fail. All the assumptions that have been used in making their forecasts have been violated by emerging real-world events. Currently (January 2021) , for example, even the professionals have no idea of how to reliably forecast what will happen to visitor arrivals in New Zealand over the next year or two. They have no real idea what the pattern of recovery will look like over what period.

There are a lot of other data series for which covid has had next to no effect (e.g., climate series), or little effect (e.g., fruit and vegetable prices as in Figure 5 below). Our current tools work fine for those.

At year 13 our students learn to work with time series that show a particular, visually-obvious, type of pattern that is quite common, a trend plus a regularly repeating seasonal component (cf. Figure 2 below).

It is important for students to see series that do look like this (and for which our tools are appropriate), and others that do not (and for which our tools are not appropriate). We recommend that formal NCEA assessments use the former type of series, but that students are shown examples of both and can distinguish between them. We give some illustrative examples after the following set of links. We conclude with descriptions of some methods that .

If you are unfamiliar with these ways of thinking and working, or if you want a reminder, see the following short videos [mins: secs]:

• [4:38]
• [6:01]
• [5:12]
• A simple but interesting extension: [5:16]

These are also on /resources/3-8/ under “Teacher Preparation” together with pdfs that contain illustrated transcripts of what the movies say.

Examples

Figure 1 below shows monthly time-series data of average numbers of overseas visitors in NZ from January 2010 until December 2020.��

Figure 1: Monthly series of average overseas visitors in NZ from January 2010 until December 2020��

Figure 2 shows a seasonal decomposition of this series using just the data up until January 2020. The thin blue line shows an estimated trend plus a repeating seasonal effect. It almost perfectly reproduces the actual data (red line). (Note: few series are this perfectly modelled by a trend plus seasonal effect). But after January 2020 covid starts to affect things, borders get closed and visitor numbers begin to fall off a cliff. The simple analysis tools we give our students cannot cope with this radical change in behaviour.��

Figure 2: Decomposition of this series from January 2010 to January 2020����

Figure 3: Forecasting visitor numbers after January 2020 using the data from 2010-January 2020����

Figure 3 uses the data up to January 2020 to forecast average visitor numbers from February 2020 on. We see the forecast (red line) with its uncertainty bands in pink. The actual data is the black line. The forecast missed what really happened by miles. Our forecasts are made by projecting forwards the patterns we have seen in the past but, in 2020, covid and border closures changed everything and the real series departed from the patterns it had followed in the past and started behaving in a way that we had never seen before.��

��Figure 4: Forecasting visitor numbers in 2021 and beyond using the data up until December 2020����

Figure 4 tries to forecast 2021 and beyond using all the data up until December 2020. We cannot trust this forecast at all because the simple (monotone) trend plus season model the Holt-Winter method is assuming is no longer applicable. The pattern has been broken and we don’t know if, when, and over what period it might “go back to normal”. Students should see this and learn a very big lesson about forecasting from it — that the world doesn’t always play nice with us and sometimes something new and unexpected happens that makes our forecasts badly wrong.

����Figure 5: Forecasting average fruit and vegetable price index values using data up until December 2020����

Figure 5 is looking at the fruit and vegetable component of the NZ consumer price index (CPI) up until December 2020 and doing a Holt-Winters forecast for 2021 and 2022. Looking at the historical data (black line) and how it is approximated by the Holt-Winters modelled trend+seasonal-effect estimates (green line) you will see that there is nothing particularly unusual about 2020. We could use this series in an assessment.����

But as we have said professionals can do better at modelling and forecasting time series than we can because they have a wider range of tools. Professional times-series analysts have methods for making adjustments for all sorts of strange behaviours.

Modelling for disrupted time series

By Rachel Passmore

[Note: this is beyond the scope of NCEA Level 3]

Disruptions to time series are a relatively common event for time series modellers to contend with. Disruptions may be caused by price changes, policy changes, definition changes and environmental changes as well as pandemics. Recovery from such disruptions can be handled in a number of ways, depending on what information about the disruption is available. For example, if a time series has been disrupted by a price change, there may have�� been other price changes in the past, so an estimate of the impact of a price change can be made and the time taken for levels to recover can be made and projections adjusted accordingly. The four most common patterns for a disruption are:

• Permanent shift to the level of the time series, either up or down.
• A ‘blip’ , perhaps caused by a one-off event; time series resumes pre-blip level and patterns afterwards.
• Permanent shift to the level of the time series but it does not happen all at once.
• Sudden shift to the level of the time series which gradually over time reverts to its pre-sudden shift level

Each of these possibilities can be incorporated into a model of time series. This part of time series analysis is known as intervention analysis and is usually modelled using an ARIMA or Autoregressive Integrated Moving Average Model although other models are possible too. Some disruptions to time series are not easy to detect, and this is where change point detection can be employed. This technique can be used to detect sudden changes in live data streams for monitoring purposes, but can also be used to detect all historic changes in the time series, not just the most obvious one

Some Australian researchers have also been developing probabilistic forecasting. They were interested in forecasting international arrivals to Australia as there is strong interest from a number of parties in knowing when pre-COVID levels might resume, if ever. Their approach was to survey 433 tourism experts and ask their opinion on pessimistic, most-likely and optimistic rates of recovery in visitor arrivals. Scenario-based��

Probabilistic forecasts were produced and compared with forecasts generated as though COVID had never happened. This allowed tourism operators to estimate the difference between the scenario-based probabilistic forecasts and the projections generated as if the pandemic had never occurred.

Intervention analysis and scenario-based probabilistic forecasting iis beyond the scope of the NZ Secondary School curriculum, but given the number of time series data sets that could be impacted by the pandemic, here are some suggestions that could provide a way forward, admittedly they are a bit of a ‘fudge’ !

1. Permanent shift to level of time series. In this instance, remove data post shift and predict from the last pre-shift data point. Make an assumption about the size of the shift, and reduce/increase projections by that amount.
2. Blip – remove this value from the data set and replace with an interpolated value, then calculate projections in the usual manner
3. Permanent shift that happens over time, such as a gradual decline. This could be be modelled by adding an additional variable in the model that decreases over time.
4. Sudden shift that gradually recovers to pre-disruption levels. This could be modelled by adding an additional variable in the model that increases over time until pre-disruption levels are regained.

The first scenario can be handled by employing iNZight in the usual way, but then adjusting projections by the estimated size of the shift. At secondary school level, this does not have to be a sophisticated estimate; as long as some justification can be provided for the size of the estimate that should be sufficient.

The second scenario just requires substitution of the one unusual data value, then iNZight can be used in the usual way.

The third scenario requires two assumptions to be made; the size of the shift, and the length of time taken for the shift in mean level to occur. This is definitely outside the scope of secondary school students in terms of assessment, but might be a useful teaching activity to discuss how this could be done. Suggestions include taking projections produced by data up to but not including any COVID effect, and then adjusting projections by applying a linear or non-linear decay element.

The fourth scenario can be handled in a similar fashion to the third scenario, but this time adjust projections by applying a growth element for a limited period of time.

References

The post Teaching Time Series in the Covid-affected Era appeared first on 91����.

Rachel’s workshop considered examples of time series data sets that can be used to promote the teaching and learning of time series. She discussed how these might differ from time series used for assessment purposes. Lots of examples of time series data sets will be provided for both purposes. Rachel��also summarised some survey data concerning teachers’ perceptions of the shift from the old time series achievement standard to the current one. Possible pedagogical responses to points made by these teachers will be discussed.

Time-series-for-teaching-learning-and-assessment

2016-time-series-handouts

LArain-time-series data -xl

The post Selecting Time Series for teaching, learning and assessing appeared first on 91����.

Students are always asking their teachers why we have to study this topic. So start the unit by answering this question first!

The primary reason that most people study time series is that they are interested in predicting the future. To do this they need to model past behaviour of a time series and hope that this pattern of behaviour will continue into the future in order to calculate a prediction. The problem is that some time series are just unpredictable, some that are not we can attempt to model.

Predictions of time series are required in many different areas

• Population projections are calculated by Government bodies in order to predict when a new school, hospital, road, bridge, prison or houses will need to be built.
• Economic forecasts, such as share prices or exchange rates are used by financial institutions.
• Weather forecasts are probably the most common types of prediction which we hear about every day.
• Environmental forecasts covering topics like global warming, monitoring populations of species close to extinction, spread of disease, rainfall, temperature etc are calculated by scientists from a variety of disciplines

As a Government Statistician I produced a variety of projections including

• Ocean wave heights for engineers who were trying to develop machines to harness wave energy and needed to know wave heights as from this they could calculate the forces that the machines would need to withstand.
• Prison population projections by type of prisoner ( lifer, remand, short-term, young offender, etc.) so that decisions about where and when to build new prisons could be taken
• Predictions of the number of prescriptions dispensed nationally. The government subsidise each prescription dispensed so from the projections could work out a budget.
• Predictions of hospital waiting lists for a variety of procedures. These predictions were not only used as part of hospital planning but also were input into medical training programmes to ensure that the right number of specialists were being trained in the right areas.

What are time series?

Provide a selection of time series for your students to discuss in groups. Some possible sources for these are

• Statistics NZ
• NZ
• 91����market
• American Statistical Association
• Google trends

Allow students to develop their own descriptions before you introduce correct terminology. Encourage students to speculate about possible reasons for the variation that they can see. Give the same time series to different groups – it is interesting to see the different features that different groups will notice. Aim to include a range of time series with increasing complexity from

• Stationary (no trend) time series with little variation
• Time series with no long term trend but some seasonality
• Time series with a linear trend
• Time series with a linear trend and seasonality
• Time series with a non-linear trend and seasonality
• Time series with a piece-wise trend , with and without seasonality
• Time series with linear trend and cycle
• Time series with a non-linear trend and cycle
• Time series with no discernible pattern i.e. one that is unpredictable

Examples of time series with some of these characteristics are given in Appendix 1. Some of the time axes are unquantified; this is deliberate and designed to stimulate debate about what the unit of measurement might be from the shape of the variation. Don’t worry too much about the meta data at this stage either, the main focus should on developing students’ skills of describing time series patterns.

Terminology

Having exposed students to a variety of time series which they described using their own vocabulary, re visit the same time series and repeat description of time series but this time using the correct terminology.

Terms to cover include:-

Trend – short term and long term. The long term trend is the most slowly changing component of the series. The trend can be either increasing or decreasing over time and it may be linear or non-linear. A short term trend is a temporary shift which may or may not have been caused by a one-off unusual event; once this event has passed the previous long term trend direction is normally resumed.

Seasonality – remind students that a ‘season’ might be a day, a week, a month, a quarter or any repeating time period.

Residuals – ask students to identify any unusual residuals, which is a residual which is greater than 10% of the overall variation in the raw data series. Any unusual values, thus identified, warrant some further investigation. Perhaps this unusual value represents an error in the data, perhaps it occurred as the result of another related unusual event; students will need to research events around the time of the unusual value to conjecture about possible reasons. Conjectures are fine, proof is not required.

Peaks and Troughs – terms used to describe local maxima and minima in a time series. Students should identify if peaks or troughs occur at the same point in the seasonal cycle and again conjecture about possible reasons for this.

Cycle – A cycle is a recurrent wave-like pattern. The period and amplitude of a cycle is neither fixed nor predictable. Thus we can describe cycles as irregular wave-like patterns in series. Many financial and economic time series have cycles that are related to changing business conditions. Students should be exposed to time series with cycles but they will not have to model them as the techniques required are far too complex.

Smoothing Techniques

Hopefully some students will have struggled to adequately describe the time series you have exposed them too, particularly if you have presented your time series with equal length axes as opposed to a longer x axis.

The overall trend is often hard to identify particularly in a series which is dominated by seasonality. In order to view just the trend without the distraction of the seasonality a number of smoothing techniques are available. If you google smoothing techniques you will see there are many. The student tutorial provided in Appendix 2 introduces a few smoothing techniques, namely

• Moving Mean
• Weighted Moving Mean
• Exponential smoothing (=0.5)
• Exponential smoothing (=0.1)

Through completion of this tutorial, students can see the effects of smoothing and this lays the foundation for using the time series module of iNZight, which uses a smoothing procedure called Holt-Winters. A Teacher’s Guide to Holt-Winter’s is attached at Appendix 3. Students do not need to know how Holt-Winters works but they do need to understand that it is a refinement of exponential smoothing, so it can be helpful to go through the process of calculating smoothed values by hand before they are exposed to the software that will handle the calculations for them. Without this step, the software becomes a ‘black-box’ and an important component of the student’s learning trajectory will have been omitted. Robust research also supports the importance of this step.

Description of overall trend

With the move to using real data in the teaching and assessment of time series, the task of describing the overall trend of the time series has become a lot more complex. No longer can scaffolding be provided by using the coefficient of a linear regression model and the thorny issue of how many pieces comprise the time series emerges. This issue is the subject of a separate paper currently being prepared by the NZSA Education Committee and will be posted on Census@School website when finalised but the advice, in short, is to train students to describe the long term trend of their time series, which means they should not be distracted by short term variations which will always be present. See examples of acceptable and unacceptable trend descriptions below.

“Looking at the smoothed values, there appears to be a slight increasing trend in the mean area of Arctic sea ice from Jan 1990 – Dec 1992, followed by a decline in the mean area of Arctic Sea Ice during 1993, a slight increasing trend in 1994, then general decreasing trend from 1995-2011(with a more rapidly decreasing trend than the rest of the years at the second half of 1995 and 2007 respectively).”

This is an unacceptable description of the trend as it focuses too much on short term variations.

“Overall the trend in the mean area of Arctic sea ice from 1990 to 2010 is slightly decreasing.”

This is an acceptable description of the overall trend; the only addition to this might be some comment quantifying the change, for example,

“The area of Arctic sea ice shows a very gradual decline over the period 1990 to 2010. The trend level has fallen from around 9.5 million km²to around 8.5 million km² over the time period.”

Predictions

If a student understands the underlying concepts of the model of their time series they can then make sensible statements about their predictions. For example, how far into the future are the predictions likely to be reliable. Are there any indications that past patterns of behaviour are not going to continue? Are predictions available for any related time series? How do these predictions compare to those calculated? What do the width of the confidence intervals tell you about the predictions? What do the width of the confidence intervals tell you about the fit of the model in general? Remember “All models are incorrect – some are useful” (Box, 1987)

Predictions are especially problematical if there has been an unusual value near the end of the time series and will be reflected in wide confidence intervals.

Model robustness can be tested by removing the last few values of a time series, re fitting the model and investigating how the ‘predictions’ compare with actual data values. If the actual data values fall within the prediction confidence intervals, model robustness is supported.

Interpretation and Conjecture

Encourage students to explain the features they have observed in their time series. Some features will be easier to explain than others, for example

Possible explanations for seasonal effects

• Ice cream sales – more sold in summer, fewer in winter
• Power usage – in NZ more power in winter, less in summer, but compare this with countries that have hotter climates. Often power usage is greater in summer because of air conditioning.
• Alcohol sales – often peaks around Christmas and New Year.
• Retail sales – again peaks around Christmas are common. Perhaps compare with countries who do not celebrate Christmas, are their seasonal patterns different?
• Weather aspects – is rainfall seasonal?

Possible explanations for changes in long term overall trend

• 2008 Global Financial Crisis. Many financial series show dramatic disruptions to overall trends around 2008.
• Global Warming. Consider related series.
• Health Scares – SARS, Asian flu, AIDS, Ebola, Mad Cow Disease. These show up well on Google trends.
• Acts of Terrorism – such acts can dramatically affect airline travel and other aspects of tourism.
• Major Sporting Events – Olympics, Commonwealth Games, Football World Cup, Rugby World Cup.
• Investigate other research to see if it confirms or refute a suspected overall trend change.

Interrogative Reasoning

At the end of an initial analysis of a time series students should consider further questions inspired by their investigation. For example, if conditions changed can they suggest how this might affect predictions? Some time series may reflect patterns shown in related time series but the pattern is lagged – i.e. the pattern in one series is several time periods behind that in another series.

In the Food for Thought data set a drop in the four retail spending series – supermarkets, fresh food, takeaways and restaurants – was found but the drop occurred in each series at different times. In this scenario a reduction in fresh food spending was the first to fall, followed by takeaways, then restaurants and supermarkets. Thus if in the future a drop is observed in fresh food spending it may be an indicator of falls to come in other related time series.

What happens next in time series analysis?

It is always good to be able to explain to students what happens next in a topic. In time series it generally means moving on to more complicated models that will enable them to model some of the time series students saw at the beginning of the unit. Models covered in a Stage 3 Time series course at the University of Auckland include:

• Autocorrelation – inclusion of an element in the model for correlation between values
• Transformation of series – for time series with a non-constant variation
• Alternative smoothing techniques
• Harmonic models – using trigonometric functions to model trends
• ARCH models – autoregressive conditional heteroscedasticity models, used to model time series whose variation changes over time.

Time Series for Teaching and Assessment

Do��use a variety of time series with linear & non-linear trends, fluctuating variation, cycles and large residuals in your teaching of time series.

Do not��use these more complex time series in your assessments. Such time series are beyond the capability of many 3^rd Year University students so don’t expect your secondary school students to cope.

This does pose a problem as real data is not often nicely behaved, yet teachers are encouraged to use real data in their assessments.

Another problem is that teachers in the conditions of assessment guidelines are requested to provide multivariate data sets for time series assessments from which students must select one time series to analyse. It is an extremely difficult and almost impossible task to find a multivariate data set with all variables providing analysis opportunities of equal difficulty. It can also represent a huge marking workload when students select different series. I suggest teachers limit the multivariate data set to 2 or 3 variables maximum.

Some teachers are also using alternative forms of assessment such as a presentation rather than a report for the time series internal in an attempt to reduce the marking workload.

The hierarchical levels of reasoning referred to in this document are taken from a framework for the development of reasoning in time series constructed for my Master’s thesis which is due to be submitted at the end of January 2016.

Rachel Passmore
November 2015

Download the Appendices

The post Suggestions for Teaching Time Series appeared first on 91����.

The post Food Spending appeared first on 91����.

1. Building consents
2. Carbon dioxide emissions in Hawaii
3. Food for thought
4. International Air passenger data
5. Polar Ice data (Updated 2017 by Marion Steel, Source: )
6. Average number of visitors in NZ by purpose
7. Dwelling consents
8. Labour Force Survey data by regions
9. Permanent & Long-term migration to & from Australia
10. Permanent & long-term migration totals
11. CPI NZ data – All Groups
12. Total passenger movements
13. Visitor arrival totals
14. Visitor departure totals
15. USA Beer production

Download data

The post 15 Time Series 91����sets (2012) appeared first on 91����.

91���� include Rainfall, NZ Alcohol consumption, Births, Travel and Tourism, Accommodation, Air Passengers and temperatures around the world.

Download data

The post Time Series 91����sets (2013) appeared first on 91����.

(Presented by Angela Hawkins in Auckland)

Resources:

• Statistics Teachers’ Day Workshop Non-sampling errors Ppt
• Evaluation of statistical reports, Teachers Booklet

Part 1: Non-sampling errors in Polls and surveys (D. Rose)

• Non-sampling errors Ppt
• Political Polls in New Zealand Ppt
• Organisation frame for poll
• Worry Questions Polls and suveys
• Writing frame critically evaluating a poll
• Attitude to Gratitude NZ herald,��article
• Kiwis unlikely to queue up for asset shares article
• Auckland toll survey
• Breaking Speed limits poll
• Broadcasting standards poll report
• Comparing polls Horizon
• DVT Com Res survey
• DVT survey Questions
• NZ-US exercises article

Part 2: Observational Studies and Experiments (R. Passmore)

• Media Articles for evaluation of statistical reports concerning experiments or observational studies
• Evaluation of reports on experiments and observational studies Ppt
• Worry Questions for evaluating media reports on observational studies ONLY

The post Non-Sampling Errors as they Apply to Polls & Surveys and Experiments & Observational Studies appeared first on 91����.

Disclaimer:

These exemplars were prepared BEFORE the final draft of the standard was made available.�� Please note that the final draft of the standard does not require a comparison of two series for MERIT level.

Download exemplars 91����set

The post Food Spending appeared first on 91����.

There will also be a discussion of why these models are an improvement over past techniques used. Details of resources that have been developed for time series analysis using iNZight will also be available.

Please note, this workshop was prepared in 2012 while the time series standard was still in draft form.���� It has since been registered and may differ slightly to the content in this workshop.

Download PowerPoint Slides

Download iNZight 91���� Files – Tips

The post Time Series Analysis using iNZight appeared first on 91����.

• Teacher’s guide to Holt Winters analysis

The post Teacher’s Guide to Holt Winters analysis appeared first on 91����.