Ltd. Data Analytics Department Analysis of the Temporal Structure in Economic Condition Assessments Copyright (c) Mizuho–DL Financial Technology Co., Ltd. All Rights Reserved.
Reserved. n This study investigates the economy watcher survey in Japan. n We elucidate the bases of assessments of economic conditions in the survey. n We aim to identify what element influences the assessments. • We focus on at which point-in-time event influences the assessments. ← Use an algorithm of learning from positive and unlabeled data (PU learning). n What we do: • Classify bases of future condition assessments into current and future events. • Find that assessments based on current events are more volatile compared to those based on future (or non-current) events. 2
All Rights Reserved. Ø Economy Watcher Survey (内閣府景気ウォッチャーデータ; Naikaku-fu Keiki Watcher Data): • A market survey published by the Japanese government. • Assessments of current and future economic conditions. Opinions of individuals from various occupations (office workers, taxi drivers, etc.) • While this survey offers crucial insights for economic policy making, the specific timing of the events that form the bases for these assessments remains unclear. • E.g. one individual might assess future economic conditions based on events that have already occurred, whereas another might base their assessment on potential future events. • Our goal is to clarify at which point-in-time events influence the assessments. 3
Co., Ltd. All Rights Reserved. n Surveys are conducted in each month and each region (Hokkaido, Tohoku, Hokuriku, …, Kyushu, Okinawa). In each month, there are about 2,500 survey responses. n In this study, we use survey data from Jan. 2016 to Dec. 2021. n In each response, there are • Assessments about current economic conditions. • Assessments about future economic conditions. n “Future”: two months later. 4 分野 景気の先 ⾏き判断 業種・職種 景気の先⾏きに対する判断理由 家計 動向 関連 (北海道) ◎ 百貨店（担当者） ・北海道新幹線の開業という外的要因があるため、今後についてはやや良くなる。 ◎ 観光名所（従業員） ・３⽉26⽇に北海道新幹線が開業することで、終着駅効果が⾒込める。 ◦ 商店街（代表者） ・⽯油製品価格が下がっているため、商材やサービスにお⾦が回ることが期待できる。 ◦ 商店街（代表者） ・３⽉は北海道新幹線の開業により、観光客を中⼼に⼈の動きが活発化することになる。駅前のホテルはすでに開業⽇前 後が満室状態であり、ホテルのほか、飲⾷関連にもプラスが⾒込める。 ◦ スーパー（役員） ・北海道新幹線の開業により、観光関連の景気が良くなり、街全体の景気の押し上げ効果が期待できる。 ◦ コンビニ（店⻑） ・客のニーズの変化に合わせた素早い対応を⾏うことで、今後も好調を維持することができる。 ◦ ⾐料品専⾨店（店⻑） ・オーダースーツなどの⾼単価のアイテムが好調で、１⽉の売上が前年⽐106％と堅調に推移していることから、この流 れは春の新作が発表される時期も続く。 ◦ 乗⽤⾞販売店（経営者） ・前年末に発売された新型⾞の受注が引き続き好調であり、今後の売上増が⾒込まれる。また、当地区の新⾞市場におい て、登録台数が全国以上の伸びを⽰しており、市場が活性化している状況がうかがえる。 ◦ 乗⽤⾞販売店（役員） ・前年の年初は消費税増税後の反動が⻑引き、販売が奮わなかったが、今年は⼤きな影響を及ぼすような要因がないため、 現状の新型⾞の好調を維持できる。 ◦ 観光型ホテル（スタッフ） ・国内外の個⼈客、団体客ともに好調であり、今後も２⽉には春節、３⽉にはプロ野球の地元開催、４⽉には⼤型学会な どがあるため、⾼単価、⾼稼働での推移が⾒込める。 ◦ 旅⾏代理店（従業員） ・流氷観光及び冬季の体験観光が充実してきたことから、２〜３⽉の外国⼈観光客の予約が好調との声が多い。外国⼈観 光客の歩留りを懸念する施設も多いが、春節の中国系の観光客に期待するところは⼤きい。
Ltd. All Rights Reserved. Ø January 2016. n Example of assessments of current economic conditions: • Home sales company (manager). • “For the past three months, the number of customers has not increased and there is a sense of stagnation.” • Score 2 (middle). 6 ◦ 観光型ホテル（スタッフ） ・中国本⼟を始めとする近隣アジアからの観光客、国内からの個⼈客、団体客のいずれも好調に推移している。需要を喚起するようなスポーツ⼤ 会イベントがあったことも追い⾵となった。 ◦ 旅⾏代理店（従業員） ・⾼額商材の受注が好調である。 ◦ タクシー運転⼿ ・好調であった12⽉の売上をそのまま維持できている。夜の会合なども増加している。 □ 商店街（代表者） ・客の財布のひもが固く、買上単価がなかなか上がってこない。 □ 商店街（代表者） ・１⽉は豪雪の影響により、道路状況が悪かったため、⾼齢者の外出機会が減少するというマイナス要因があった。ただ、全体的なトレンドには 変化がみられない。 □ 商店街（代表者） ・客の声としては、変化が感じられないという意⾒が⼤半である。 □ 商店街（代表者） ・１⽉に⼊り、プレミアム付商品券の効果が⼀部の業種や商店街内の百貨店の⼀部店舗などに表れている。ただし、プレミアム付商品券の使⽤期 間が１⽉までであるにもかかわらず、売上の伸びは鈍重であった。また、12⽉に売上を伸ばした夜型飲⾷店は、客の夜間の出控えもあり、今⼀つ であった。当商店街内の客⾜も例年より途絶えがちであった。 □ 商店街（代表者） ・新年の初売りは好スタートだったが、曜⽇並びの関係から前年を下回ったところが多かった。また、積雪の少ない状況が続いていることから、 客の外出機会の増加に伴う売上増を期待したが、氷点下の⽇が多かったこともあり、伸び悩んだ。当地のプレミアム付商品券は１⽉末までが使⽤ 期間であり、これまで順調に使われているため、今後の動向に不安がある。
Ltd. All Rights Reserved. Ø January 2016. n Example of assessments of future economic conditions: • Employee in a tourist spot. • “With the opening of the Hokkaido Shinkansen (bullet train) on March 26, the terminal station effect is expected.” • Score 4 (better). 7 ◦ タクシー運転⼿ ・冬のイベントが多数組まれていることで、観光レジャー客の動きが期待できるため、今後についてはやや良く なる。 ◦ 美容室（経営者） ・１⽉はあまり良くなかったが、この状態が続くようには感じられないため、１⽉と⽐較すればやや良くなる。 ◦ その他サービスの動向を把握できる者 ［フェリー］（従業員） ・春の訪れとともに観光客の増加が期待できる。 ◦ 住宅販売会社（経営者） ・原油安により灯油やガソリンなどのコストダウンも⼗分図られてきているため、株式市場が落ち着きを取り戻 せば、来春に予定されている消費税増税を控えて、駆け込み需要が少しずつ出てくることになり、今よりも景気 はやや良くなる。 □ 商店街（代表者） ・社会情勢などが混沌としているため、客も将来がどうなるのか分からないという状況にあり、今後も変わらな いまま推移する。 □ 商店街（代表者） ・４⽉中旬までは来街者の増加が⾒込める要素はない。４⽉下旬になれば、ゴールデンウィーク前ということも あり、近隣住⺠による買物が若⼲増加するが、４⽉の外国⼈観光客の⼊込が前年よりも減少する兆候もみられる ため、景気は悪いまま推移する。 □ 商店街（代表者） ・外国⼈観光客の購⼊率が⼀服し始めていることに加えて、国内観光客による売上も少しずつ落ちてきているた め、今後も変わらないまま推移する。
Technology Co., Ltd. All Rights Reserved. n As explained, there are two types of assessments: • Assessments about the current and future economic conditions. n Question: What are bases of the assessments about the future economic conditions? • On what point-in-time events are the respondents basing their assessments of future economic condition? • Based on current events? Because now is …, the future condition will be good. • Based on potential future events? Because … will be occurred in the future, the future condition will be good. 8
Technology Co., Ltd. All Rights Reserved. n To explain the idea of the temporal structure, we raise two examples. n Example 1: Survey in October. • The current condition (October) is bad because of the current inflation. • The future condition (December) will be bad because of the current inflation. • The current inflation is a current event. 9 Current inflation
Technology Co., Ltd. All Rights Reserved. n Example 2: Survey in October. • The current condition (October) is bad because of the current inflation. • The future condition (December) will be good because of the Christmas business. • The current inflation is a current event. • The Christmas business is a potential future event. • Assessments are based on different point—in-time events. 10 Current inflation Christmas business
Technology Co., Ltd. All Rights Reserved. n Policy maker “How do we interpret the assessments in the survey?” • Some people may answer that the economic condition in December will be good. ↔ It is just because of the Christmas event. • Other people think that both current and future economic conditions are bad because of the inflation. n Policy maker might become optimistic about the future condition from the first result. ↔ This improvement of the assessments in December might be temporary. Even if the condition seems to be improved, the policy maker should not believe it. Ø Understanding temporal structure will contribute to the policy making. 11
Ltd. All Rights Reserved. n Understanding the temporal context in which people base their assessments is vital. n Goal of analysis: • Classify assessments of future economic conditions into assessments based on current and future events. • For classification, • Define current and future events. • Employ an algorithm for learning from positive and unlabeled data (PU learning). 12
Rights Reserved. n We define current and future events Ø Assumption (Definition of current event = definition of positive): All assessments of the current economic condition is based on current events. • This assumption defines what a current event is. → Future events are defined as “non-current events.” = Events that are not used to base the current economic conditions. 13
Rights Reserved. n Apply the binary classification framework. • Classify data into positive and negative data. • Positive data = current events. Negative data = future events. n Standard learning: classification using observed positive and negative data n PU learning: claassification using observed positive and unlabeled data • Unlabeled data: mixture of positive and negative data. • We can interpret that the method classifies positive and non-positive data. • Once we define “positive (current)”, we can define “non-positive (non-current)”. 14
Rights Reserved. n Embedding: transform a document 𝑆 into a 𝑑-dimensional vector 𝑋 ∈ ℝ!. n Consider distributions of 𝑋 of positive 𝑝(𝑥 ∣ 𝑦 = +1), negative 𝑝(𝑥 ∣ 𝑦 = −1), and unlabeled data 𝑝(𝑥). n Here is an idea of PU learning. The problem is that 𝑝(𝑥 ∣ 𝑦 = −1) cannot be observed directly. 15 Positive data distribution 𝑝(𝑥 ∣ 𝑦 = +1) Unlabeled data distribution 𝑝(𝑥) = Mixture of positive and negative data distributions. Scaling Scaled positive data distribution 𝑝 𝑦 = +1 𝑝(𝑥 ∣ 𝑦 = +1) − Negative data distribution 𝑝(𝑥 ∣ 𝑦 = −1) = Non-positive data distribution 𝑝(𝑦 = −1)𝑝 𝑥 𝑦 = −1 = 𝑝 𝑥 − 𝑝 𝑦 = +1 𝑝(𝑥 ∣ 𝑦 = +1)
Rights Reserved. • Features of current and future economic conditions follow 𝑝(𝑥 ∣ 𝑦 = +1) and 𝑝(𝑥). • We have a pair of observations: Positive data (embedding of current assessments): 𝑋/ /01 - ∼ 𝑝(𝑥 ∣ 𝑦 = +1). Unlabeled data (embedding of future assessments): 3 𝑋2 201 3 ∼ 𝑝(𝑥). • Train a classifier 𝑓: ℝ+ → ℝ using the positive 𝑋/ /01 - and unlabeled 3 𝑋2 201 3 datasets. • Estimate a label of 𝑥 as 7 𝑦 ≔ sgn(𝑓(𝑥)). • In some case, 𝑓(𝑥) is a model of 𝑝(𝑦 = +1 ∣ 𝑥). 17
Rights Reserved. • Zero-one loss: ℓ 𝑧 ≔ − 1 4 sgn 𝑧 + 1 4 . In practice, we replace it with some surrogate loss. n PU Classification risk: 𝑅 𝑓 ≔ 𝜋𝔼, ℓ 𝑓 𝑋 − 𝜋𝔼, ℓ −𝑓 𝑋 + 𝔼. ℓ −𝑓 𝑋 • This is equivalent to the standard binary classification risk (du Plessis et al., 2015). −𝜋𝔼+ ℓ −𝑓 𝑋 + 𝔼, ℓ −𝑓 𝑋 = 𝔼- ℓ −𝑓 𝑋 . • We replace 𝔼, and 𝔼. by the sample averages of 𝑋/ /01 - and 3 𝑋2 201 3 . • We train a classifier by minimizing the empirical risk. 18 Risk of misclassifying positive data as negative Risk of misclassifying negative data as positive − =
All Rights Reserved. Ø We model a classifier 𝑓 by neural networks. n We should train different neural networks for each month because the contents (words, etc) of assessments change across months (distribution shift). → However, using a different model in each month is inefficient in training a model. n Multi-task learning framework: • Share a part of the structure of neural networks over all periods. 19
Rights Reserved. n We trained neural networks using data from Jan. 2016 to Dec. 2019. n We plot the average original and average estimated scores of economic conditions. • The Green and Red dot lines are average original scores. • The Brue and Orange lines are our average estimated scores. • Scores about conditions: 0 (worse), 1, 2 (no change), 3, 4 (better). 20
Rights Reserved. n Green dot line: average scores of original current economic condition n Red dot line: average scores of original future economic conditions. • Current conditions have higher average scores than future conditions. 21
Rights Reserved. n Brue line: average estimated scores of future condition based on current events. n Orange line: average estimated scores of future condition based on future events. • The Brue line is more volatile than the Orange line. → Current events are changeable. Future events are affected by fundamental elements. 22
Technology Co., Ltd. All Rights Reserved. n Co-occurrence network of assessments based on current events in June 2016. • “Kumamoto earthquakes.” （April 2016） • “May” and “June.” • “Chugen (中元).” • “rainy season” → They are related to events occurred this or recent months. 23
Technology Co., Ltd. All Rights Reserved. Ø Co-occurrence network of assessments based on future events in June 2016. • “Brexit referendum” (23 June 2016). • “trend” and “business cycle.” • “tax increase.” • ”Bonus.” n “trend” and “business cycle” represent economic fundamentals. • Unlike current economic conditions, people use fundamentals to predict future economic conditions. 24
Rights Reserved. 25 n Brue line: average estimated scores of future condition based on current events. n Orange line: average estimated scores of future condition based on future events.
Technology Co., Ltd. All Rights Reserved. n Co-occurrence network of assessments based on current events in Feb. 2017. • “Valentine’s day.” • “Chinese new year.” • “January” and “February.” 26
Technology Co., Ltd. All Rights Reserved. n Co-occurrence network of assessments based on future events in Feb. 2017. • “consumption” and “recovery.” • “construction.” • “USA president.” n “consumption” and “recovery” represent economic fundamentals. • People assess the future economic conditions based on events that will occur or unchangeable matters (fundamentals). 27
Reserved. Ø Assessments of future economic conditions: classified them based on their bases. n Assessments based on current events: • Events that have already occurred by the month. n Assessments based on future events. • Events that are expected to occur. • Economic fundamentals (consumption, construction, tax, etc.). It might be hard to interpret the fundamentals as “future” events if we use “future” in a daily way. However, our future events are defined as bases that are not used in the assessments of current economic conditions. → Economic fundamentals can be interpreted as non-current events (= future events) in this study from the definition. • Fundamentals do not change → keep the assessments less volatile. 28
Reserved. n We classified assessments abut future economic conditions into those based on current events and future events. n We defined current events as bases that are used to assess current economic conditions and future events as non-current events. n We employed the PU learning method for classification. n We find that • We can classify the assessments. • Scores for assessments based on future events are less volatile. 29
masakat0
ontheslope
kikuzo
koide3
reokashiwa
skydats
gitomochang
mickey_0226
yukinobaba
ayaniwa
kohei_1979
.png){kind=avatar}
satai
mfonobong
skipperchong
tammyeverts
akashhashmi
sabderemane
.png){kind=avatar}
helenjbeal
vanstee
portentint
ianozsvald
samtorres
cassininazir
baasie
