Upgrade to Pro — share decks privately, control downloads, hide ads and more …

Prophetを使った時系列予測

 Prophetを使った時系列予測

- Prophetアルゴリズムの基本
- 検証のやり方
- 季節性での加法と乗法の使い分け
- 変数重要度と変数ごとの効果

19fc8f6113c5c3d86e6176362ff29479?s=128

Kan Nishida
PRO

February 26, 2020
Tweet

Transcript

  1. Exploratory ηϛφʔ #25 ProphetΛ࢖ͬͨ࣌ܥྻ༧ଌ

  2. EXPLORATORY

  3. 3 εϐʔΧʔ ੢ా צҰ࿠ CEO EXPLORATORY ུྺ 2016೥य़ɺσʔλαΠΤϯεͷຽओԽͷͨΊɺExploratory, Inc Λཱ

    ্ͪ͛Δɻ Exploratory, Inc.ͰCEOΛ຿ΊΔ͔ͨΘΒɺσʔλαΠΤϯεɾϒʔ τΩϟϯϓɾτϨʔχϯάͳͲΛ௨ͯ͠σʔλαΠΤϯεͷٕज़ͱख ๏ͷීٴͱڭҭʹऔΓ૊Ήɻ ถΦϥΫϧຊࣾͰɺ16೥ʹΘͨΓσʔλαΠΤϯεͷ։ൃνʔϜΛ཰ ͍ɺػցֶशɺϏοάɾσʔλɺϏδωεɾΠϯςϦδΣϯεɺσʔ λϕʔεʹؔ͢Δ਺ଟ͘ͷ੡඼ΛੈʹૹΓग़ͨ͠ɻ @KanAugust
  4. Vision ΑΓΑ͍ҙࢥܾఆΛ͢ΔͨΊʹ σʔλΛ࢖͏͜ͱ͕౰ͨΓલʹͳΔ

  5. Mission σʔλαΠΤϯεͷຽओԽ

  6. 6 ୈ̏ͷ೾ σʔλαΠΤϯεɺAIɺػցֶश͸౷ܭֶऀɺ։ൃऀͷͨΊ͚ͩͷ΋ͷͰ͸͋Γ·ͤΜɻ σʔλʹڵຯͷ͋ΔਓͳΒ୭΋͕ੈքͰ࠷ઌ୺ͷΞϧΰϦζϜΛ࢖ͬͯ ϏδωεσʔλΛ؆୯ʹ෼ੳͰ͖Δ΂͖Ͱ͢ɻ Exploratory͕ͦ͏ͨ͠ੈքΛՄೳʹ͠·͢ɻ

  7. ୈ1ͷ೾ ୈ̎ͷ೾ ୈ̏ͷ೾ ϓϥΠϕʔτ(ߴ͍/ݹ͍) Φʔϓϯɾιʔε(ແྉ/࠷ઌ୺) UI & ϓϩάϥϛϯά ϓϩάϥϛϯά 2016

    2000 1976 ϚωλΠθʔγϣϯ ίϞσΟςΟԽ ຽओԽ ౷ܭֶऀ σʔλαΠΤϯςΟετ Exploratory ΞϧΰϦζϜ Ϣʔβʔɾ ମݧ πʔϧ Φʔϓϯɾιʔε(ແྉ/࠷ઌ୺) UI & ࣗಈԽ ϏδωεɾϢʔβʔ ςʔϚ σʔλαΠΤϯεͷຽओԽ
  8. 質問 ExploratoryɹϞμϯˍγϯϓϧ UI 伝える データアクセス データ ラングリング 可視化 アナリティクス 統計/機械学習

  9. Exploratory ηϛφʔ #25 ProphetΛ࢖ͬͨ࣌ܥྻ༧ଌ

  10. ࣌ܥྻ༧ଌϞσϧ • ڭࢣ͋ΓֶशͷҰछɻ • աڈͷ࣌ܥྻσʔλΛτϨʔχϯάσʔλͱͯ͠ɺকདྷͷ஋Λ༧ଌ ͢ΔϞσϧΛ࡞Δɻ 10

  11. 11 աڈN೔ͷσʔλ͔Βɺ࣍ͷ೔ͷ஋Λ༧ଌ͢ΔϞσϧΛ࡞੒͢Δɻ ࣌ܥྻ༧ଌ - ఻౷తͳΞϓϩʔν Day 1 Day 2 Day

    3 Day 4 Day 2 Day 3 Day 4 Day 5 Day 3 Day 4 Day 5 Day 6
  12. 12 • σʔλؒͷִ࣌ؒؒ͸σʔλΛ௨ͯ͠ҰఆͰ͋Δඞཁ͕͋Δɻ • ஋͕NA (ܽଛ஋)ͱͳΔ೔͕͋ͬͯ͸ͳΒͳ͍ɻ • ෳ਺ͷपظੑ (िͱ೥) Λಉ࣌ʹѻ͏ͷ͸೉͍͠ɻ

    • ύϥϝʔλͷઃఆʹɺઐ໳తͳ஌͕ࣝඞཁɻ ఻౷తͳΞϓϩʔνͰͷ࣌ܥྻ༧ଌͷ໰୊
  13. 13 • Facebookʹ͍ͨσʔλαΠΤϯςΟετ ʢSean J. Taylor & co.ʣ͕࡞ͬͨ࣌ܥྻ༧ ଌΞϧΰϦζϜͰɺΦʔϓϯιʔεͱͯ͠ ެ։͞Ε͍ͯΔɻ(https://

    facebook.github.io/prophet) • ౷ܭɺ࣌ܥྻ༧ଌͷઐ໳஌͕ࣝͳͯ͘΋ ࢖͑ΔΑ͏ʹσβΠϯ͞Ε͍ͯΔɻ Prophet Sean J. Taylor @seanjtaylor
  14. 14 • ҎԼͷཁૉͷ࿨ͱͯ͠ද͢͜ͱͷͰ͖Δ׈Β͔ͳۂઢͷ͏ͪɺաڈ σʔλʹ࠷΋ϑΟοτ͢Δ΋ͷΛ୳͢ɻͦͷۂઢΛະདྷʹԆ௕͢Δ ͜ͱʹΑͬͯ༧ଌ͢Δɻ • େہతͳ੒௕τϨϯυ • पظͷقઅੑ (೥ɺिɺ೔ͳͲ)

    • ॕ೔ޮՌ - ΫϦεϚεɺ৽೥ɺ೥࣍ΠϕϯτɺͳͲɻ • ֎෦༧ଌม਺ Prophet
  15. 15 େہతͳτϨϯυ

  16. 16 େہతͳτϨϯυ

  17. 17 ೥पظͷقઅੑ

  18. 18 τϨϯυ + ೥पظੑ

  19. 19 िपظͷมಈ

  20. 20 τϨϯυ + ೥पظੑ + िपظੑ

  21. 21 τϨϯυ + ೥पظੑ + िपظੑ 21

  22. 22 ॕ೔ޮՌ

  23. 23 τϨϯυ + ೥पظੑ + िपظੑ + ॕ೔ޮՌ

  24. 24 • ࣌ؒͷൃలΛϞσϧͰදݱ͢Δ͜ͱ͸͖͋ΒΊΔɻ • ͔ΘΓʹɺ୯ʹۂઢΛݟ͚ͭΔͱ͍͏໰୊ʹ͢Δ͜ͱʹΑͬͯҎԼͷΑ͏ͳ ར఺Λಘ͍ͯΔɻ • σʔλؒͷִ͕࣌ؒؒҰఆͰ͋Δඞཁ͸ͳ͍ɻ • ஋͕NA

    (ܽଛ஋)ͱͳΔ೔͕͋ͬͯ΋໰୊ͳ͍ɻ • ෳ਺ͷपظੑ (िͱ೥) ͕σϑΥϧτͰߟྀ͞ΕΔɻ • σϑΥϧτͷઃఆͰͦΕͳΓͷ༧ଌ͕Ͱ͖ΔɻઃఆՄೳͳύϥϝʔλͷଟ ͘͸ઐ໳஌ࣝແ͠ͰཧղՄೳɻ Prophetͷར఺
  25. Let’s do it! 25

  26. Global Sales σʔλ 26

  27. 27 Sales (ച্)ΛϥΠϯνϟʔτͰՄࢹԽͯ͠ΈΔ • X࣠ʹOrder DateΛɺ࣌ؒ୯ҐΛ WEEKͱׂͯ͠Γ౰ͯΔɻ • Y࣠ʹSalesΛɺؔ਺ΛSUM(߹ ܭ)ͱׂͯ͠Γ౰ͯΔɻ

  28. 28

  29. Salesʢച্ʣΛ༧ଌ͢ΔϞσϧΛ࡞Δ 29

  30. ΞφϦςΟΫεɾϏϡʔΛબ୒͠ɺ࣌ܥྻ༧ଌΛબ୒ɻ 30

  31. • ೔෇/࣌ؒͷྻʹOrder DateΛ࣌ ؒ୯ҐΛWEEKͱׂͯ͠Γ౰ͯ Δɻ • ஋ͷྻʹSalesΛɺؔ਺Λ SUM(߹ܭ)ͱׂͯ͠Γ౰ͯΔɻ 31

  32. ϓϩύςΟɾμΠΞϩάΛ࢖ͬͯɺ 1೥ؒͷ༧ଌΛ͢ΔͨΊʹɺ༧ଌظ ؒΛ52 (1೥ʹ52ि) ʹઃఆ͢Δɻ 32

  33. ࣮ߦϘλϯΛΫϦοΫ͢Δͱ༧ଌ͞Εͨσʔλ͕ΦϨϯδ৭ͷϥΠϯͰදࣔ͞ΕΔɻ 33 ΦϨϯδ৭ͷઢ͚ͩͷ۠ؒ͸ࠓޙ1೥ؒͷ༧ଌΛද͢ɻ
 ୶͍ΦϨϯδ৭͸༧ଌͷෆ֬ఆ۠ؒ(uncertainty interval)Λද͢ɻ

  34. 34 τϨϯυ λϒΛΫϦοΫ͢ΔͱτϨϯυϥΠϯͷೖͬͨνϟʔτ͕දࣔ͞ΕΔ άϦʔϯͷઢ͕τϨϯυϥΠϯ

  35. 35 ྘ͷόʔ͸ɺτϨϯυʹมԽ͕͋ͬͨ࣌఺ʢνΣϯδϙΠϯτʣΛද͢ɻ ๮ͷߴ͞͸ɺνΣϯδϙΠϯτͰͷτϨϯυϥΠϯͷ܏͖ͷมԽྔɻ

  36. 36 ೥पظλϒΛΫϦοΫ͢Δͱ೥पظͷνϟʔτ͕දࣔ͞ΕΔɻ

  37. Every year, the sales doesn’t pick up until June, then

    it goes down in July.
  38. Weekly tab shows up only when the data is daily

    or more granular levels (hour, minutes, etc.)
  39. Every week, the sales are low on Sunday and Monday,

    and the rest of the week is high.
  40. 40 िपظΛݟΔͨΊʹɺ࣌ؒ୯ҐΛ೔ʹมߋ࣮ͯ͠ߦ͠௚͢ɻ

  41. σʔλͷલॲཧ ܽଛ஋ʢNAʣͷॲཧ

  42. Somehow, the weekly seasonality doesn’t repeat exactly the same…

  43. Somehow, the weekly seasonality doesn’t repeat exactly the same…

  44. There are NA for some dates. You can impute NA

    as part of the Data Preprocessing.
  45. 45 • ஫จ͕Ұ݅΋ͳ͍೔ͷച্Λ0ͱͯ͠ѻ͏ͨ Ίɺ஋ͷྻʹ஋͕ͳ͍ͱ͖ͷॲཧʹ”θϩͰ ຒΊΔ”ΛબͿɻ • ͜ΕΛ͠ͳ͍৔߹ɺ஫จ͕Ұ݅΋ͳ͍೔ ͸ɺσʔλ͕ແ͍೔ͱͯ͠ѻΘΕΔɻ

  46. None
  47. None
  48. Under the Importance tab, you can see which seasonality has

    more effect on the forecasting outcome.
  49. 49 िपظλϒΛΫϦοΫ͢Δͱिपظͷνϟʔτ͕දࣔ͞ΕΔɻ

  50. 50 ޮՌλϒΛΫϦοΫ͢Δͱɺ༧ଌ஋Λߏ੒͢Δ֤ཁૉʢτϨϯυɺقઅੑʣ͕දࣔ͞ΕΔɻ

  51. 51 ม਺ॏཁ౓λϒΛΫϦοΫ͢ΔͱԿ͕༧ଌʹେ͖ͳӨڹΛ༩͍͑ͯΔͷ͔͕ݟ͑Δɻ

  52. 52 σʔλλϒΛΫϦοΫ͢Δͱ༧ଌ෇͖ͷσʔλ͕දࣔ͞ΕΔ

  53. • forecasted_value - ༧ଌ஋ • forecasted_value_high/forecasted_value_low - ෆ֬ఆ۠ؒ • trend

    - େہతͳ੒௕τϨϯυ • yearly - ೥पظͷτϨϯυ • weekly - िपظͷτϨϯυ 53 ༧ଌ෇͖ͷσʔλͷಡΈํ
  54. 54 ࣌ܥྻ༧ଌͷධՁ

  55. 55 όοΫςετ • աڈσʔλͷ͏ͪɺ৽͍͠ظؒΛςετ༻ʹͱ͓ͬͯ͘ɻ • ςετظؒͷσʔλΛɺͦΕΑΓҎલͷσʔλΛ࢖ͬͯ༧ଌ͢Δ • ༧ଌσʔλͱɺ࣮ࡍͷςετظؒͷσʔλΛൺֱධՁ͢Δɻ ࣌ܥྻ༧ଌͷςετ๏

  56. σʔλΛ2ͭͷηΫγϣϯʹ෼͚Δɻ 56 τϨʔχϯάظؒ ςετظؒ

  57. 57 τϨʔχϯάσʔλΛ࢖ͬͯ༧ଌϞσϧΛ࡞ΓɺςετظؒΛ༧ଌɻ ςετظؒͷ࣮ଌ஋ͱ༧ଌ஋ͷʮͣΕʯͷେ͖͞ΛධՁ͢Δɻ τϨʔχϯάظؒ ςετظؒ

  58. 58 • ࣌ؒͷ୯ҐʹMON(݄)Λࢦఆ͢Δɻ ςετϞʔυʹ੾Γସ͑Δ

  59. 59 ςετϞʔυΛTRUEʹ͠ɺςετظؒΛ12 (݄) ͱ͢Δɻ

  60. ࠨଆͷ੨͍ઢ͸τϨʔχϯάσʔλɺӈଆͷਫ৭ͷઢ͸ςετ σʔλɻ 60

  61. 61 ੨৭ͷઢ͕τϨʔχϯάσʔλɻ ͜ͷൣғσʔλΛ΋ͱʹɺ༧ଌϞσϧ͕࡞ΒΕΔɻ

  62. ΦϨϯδͷઢ͕ɺτϨʔχϯάσʔλΛݩʹ࡞ΒΕͨϞσϧʹ ΑΔ༧ଌσʔλɺബ͍ΦϨϯδͷྖҬ͸ͦͷ৴པ۠ؒͰ͋Δɻ 62

  63. ͜ͷൣғͷσʔλΛ΋ͱʹ༧ଌϞσϧ͕࡞ΒΕ͍ͯΔͨΊɺ
 ͜ͷൣғͷ࣮σʔλͱϞσϧʹΑΔ༧ଌσʔλ͸͔ͳΓҰக͍ͯ͠Δɻ 63

  64. ςετظؒͷ༧ଌσʔλͱ࣮ଌ஋ͷζϨ͕ͲΕ͚ͩେ͖͍ͷ͔Λ ࢉग़͠ɺͦΕΛ࣋ͬͯϞσϧͷਫ਼౓ΛධՁ͍ͨ͠ɻ 64

  65. ςετ݁ՌͷαϚϦ 65

  66. 66 • RMSE (Root Mean Square Error) : ༧ଌ͔ΒͷͣΕͷೋ৐ͷฏۉͷϧʔτ •

    MAE (Mean Absolute Error) : ༧ଌ͔ΒͷͣΕͷઈର஋ͷฏۉ • MAPE (Mean Absolute Percentage Error) : ύʔηϯτͰදͨ͠༧ଌ͔Βͷ ͣΕͷઈର஋ͷฏۉ • MASE (Mean Absolute Scaled Error) : MAEΛɺτϨʔχϯάσʔλͰͷφ Πʔϒ༧ଌʢҰͭલͷظͱಉ͡஋͕ݱΕΔͰ͋Ζ͏ͱ͍͏୯७ͳ༧ଌʣ ͷMAEͰׂͬͨ΋ͷɻ ࣌ܥྻ༧ଌͷධՁࢦඪ
  67. Rootʢฏํࠜʣ
 Meanʢฏۉʣ
 Squareʢ2৐ʣ
 Errorʢޡࠩʣ ͭ·Γɺ࣮ଌ஋ͱ༧ଌ஋ͷޡࠩ Λ2৐ͯ͠ɺͦͷฏۉΛͱΓɺͦ ͷ஋ͷฏํࠜΛͱͬͨ஋ͷ͜ ͱɻ 67 RMSE

    (Root Mean Square Error)
  68. 
 22 + 22 + 22 + 42 
 4

    (఺ͷ਺) 4 + 4 + 4 + 16 
 4 7 = 2.65 68 RMSE (Root Mean Square Error) 2 2 4 2 = = ྫ͑͹ɺ࣮ଌ஋ͱ༧ଌ஋ͷޡ͕ࠩͦΕ ͧΕ2, 2, 2, 4ͩͬͨͱ͢Δͱɺܭࢉ͸ ҎԼͷΑ͏ʹͳΔɻ
  69. Meanʢฏۉʣ
 Absoluteʢઈର஋ʣ
 Errorʢޡࠩʣ ͭ·Γɺ࣮ଌ஋ͱ༧ଌ஋ͷޡࠩ ͷઈର஋ͷฏۉ͜ͱɻ 69 MAE (Mean Absolute Error)

  70. 
 2 + 2 + 2 + 4 
 4

    (఺ͷ਺) 70 ྫ͑͹ɺ࣮ଌ஋ͱ༧ଌ஋ͷޡ͕ࠩ ͦΕͧΕ2, 2, 2, 4ͩͬͨͱ͢Δ ͱɺܭࢉ͸ҎԼͷΑ͏ʹͳΔɻ = 2.5 MAE (Mean Absolute Error) 2 2 4 2
  71. Meanʢฏۉʣ
 Absoluteʢઈର஋ʣ
 Percentageʢׂ߹ʣ
 Errorʢޡࠩʣ ͭ·Γɺ࣮ଌ஋ͱ༧ଌ஋ͷޡࠩ ͷׂ߹ͷઈର஋ͷฏۉ͜ͱɻ 71 MAPE (Mean Absolute

    Percentage Error)
  72. 72 12 13 16 11 ·ͣɺ࣮ଌ஋Λ΋ͱΊΔɻ MAPE (Mean Absolute Percentage

    Error)
  73. 73 12 13 16 11 MAPE (Mean Absolute Percentage Error)

    2 2 4 2 ࣍ʹɺ࣮ଌ஋ͱ༧ଌ஋ͷޡࠩ Λ΋ͱΊΔɻ
  74. 74 100 100 100 100 MAPE (Mean Absolute Percentage Error)

    16.6% 15.4% 25% 18.2% ࣮ଌ஋ͱ༧ଌ஋ͷޡࠩΛ࣮ଌ஋Ͱ ׂͬͯ100Λ͔͚ɺͦΕͧΕͷ ύʔηϯςʔδΛ΋ͱΊΔɻ ਺ࣈ͕ϚΠφεͷ৔߹ɺϚΠφε ͷූ߸ΛͱΔ (ઈର஋).
  75. 75 100 100 100 100 MAPE (Mean Absolute Percentage Error)

    16.6% 15.4% 25% 18.2% 
 16.6 + 15.4 + 18.2 + 25 
 4 (఺ͷ਺) ࠷ޙʹɺ͜ΕΒͷ஋ͷฏۉΛग़͢ɻ = 18.8%
  76. 76 Meanʢฏۉʣ
 Absoluteʢઈର஋ʣ
 Scaledʢεέʔϧௐ੔ࡁΈͷʣ
 Errorʢޡࠩʣ MAEΛҟͳΔεέʔϧͷσʔλ Ͱͷ༧ଌͲ͏͠Ͱ΋ൺֱՄೳͳ Α͏ʹεέʔϧௐ੔ͨ͠΋ͷɻ MASE (Mean

    Absolute Scaled Error)
  77. 77 MASE = ςετظؒͷMAE / τϨʔχϯάظؒͷφΠʔϒ༧ଌͷMAE MASE ςετظؒͷMAE τϨʔχϯάظؒͷφΠʔϒ༧ଌͷMAE

  78. 78 Ұظલͷ஋͕ɺࠓظ΋ͦͷ··ग़ΔͩΖ͏ɺͱ͍͏҆қͳ༧ଌɻ φΠʔϒ༧ଌ

  79. 79 φΠʔϒ༧ଌΛτϨʔχϯάظؒʹରͯ͠ߦͬͨͱ͖ͷMAEΛج४ͱͯ͠࠾༻ MASEͷεέʔϧௐ੔ͷج४

  80. 80 MASE = ςετظؒͷMAE / τϨʔχϯάظؒͷφΠʔϒ༧ଌͷMAE MASE ςετظؒͷMAE τϨʔχϯάظؒͷφΠʔϒ༧ଌͷMAE

  81. 81 قઅੑͷϞʔυ Ճ๏త vs ৐๏త

  82. 82 The difference between the actual line and the forecasted

    line becomes wider as the time progresses.
  83. Կ͕ى͖͍ͯΔͷ͔ʁ 83 • ച্͕੒௕͢ΔʹͭΕɺͦΕʹ͋ΘͤͯقઅੑʹΑΔมಈ΋େ͖͘ͳΔͱ ߟ͑Δͷ͕ࣗવɻ • ͔͠͠Ϟσϧ͸قઅੑʹΑΔมಈͷେ͖͞͸͍ͭͰ΋ҰఆͰ͋Δͱ͍͏લ ఏͰ༧ଌ͍ͯ͠Δɻ • ͕࣌ؒͨͬͯച্͕੒௕ͨ͋͠ͱͰ͸ɺϞσϧ͕༧ଌ͢Δقઅมಈͷେ͖

    ͕͞ɺ࣮ࡍͷ஋ͷقઅมಈͷେ͖͞ʹ͍͍͚͍ͭͯͯͳ͍ͷͰ͸ͳ͍͔ʁ
  84. ࿨ͱͯ͠੒Γཱ͍ͬͯΔ஋ͷϦχΞͳ৳ͼํ ੵͱͯ͠੒Γཱ͍ͬͯΔ஋ͷෳརޮՌͷ͋Δ৳ͼํ མͪண͍ͯΔձࣾͷैۀһ਺ͷਪҠ Amazonͷച্ߴͷਪҠ ੒௕͢Δͱ͖ͷ஋ͷ৳ͼํͷҧ͍ 84

  85. ઌఔͷςετϞʔυͰͷ༧ଌ݁ՌΛ΋͏Ұ౓ݟͯΈΔͱ… 85 ࣮ࡍͷقઅมಈ͸ɺ༧ଌ݁ՌΑΓ΋ େ͖͘ͳ͍ͬͯΔΑ͏ͩɻ

  86. Կ͕ى͖͍ͯΔͷ͔ʁ 86 ച্ͷ੒௕ʹͱ΋ͳͬͯେ͖͘ͳΔقઅมಈΛϞσϧԽͰ͖Ε͹Α͍ͷͰ͸ʁ • ച্͕੒௕͢ΔʹͭΕɺͦΕʹ͋ΘͤͯقઅੑʹΑΔมಈ΋େ͖͘ͳΔͱ ߟ͑Δͷ͕ࣗવɻ • ͔͠͠Ϟσϧ͸قઅੑʹΑΔมಈͷେ͖͞͸͍ͭͰ΋ҰఆͰ͋Δͱ͍͏લ ఏͰ༧ଌ͍ͯ͠Δɻ •

    ͕࣌ؒͨͬͯച্͕੒௕ͨ͋͠ͱͰ͸ɺϞσϧ͕༧ଌ͢Δقઅมಈͷେ͖ ͕͞ɺ࣮ࡍͷ஋ͷقઅมಈͷେ͖͞ʹ͍͍͚͍ͭͯͯͳ͍ͷͰ͸ͳ͍͔ʁ
  87. قઅੑϞʔυ 87 Ճ๏త ଍͠ࢉͰޮՌ͕ݱΕΔɻ
 ྫɿ12݄͸ϓϥε$100,000
 ΋ͱͷ஋ʹؔΘΒͣมಈͷେ͖͞͸Ұ ఆɻ ৐๏త ֻ͚ࢉͰޮՌ͕ݱΕΔ
 ྫɿ12݄͸ϓϥε10%


    ΋ͱͷ஋͕େ͖͚Ε͹มಈ΋େ͖ ͘ͳΔɻ
  88. 88 Ճ๏త ৐๏త

  89. 89 Ճ๏త ৐๏త قઅੑͷେ͖͞͸Ұఆ

  90. 90 Ճ๏త ৐๏త قઅੑͷେ͖͞͸Ұఆ قઅੑ͸ݩͷ஋ʢτϨϯυʣʹൺྫͯ͠େ͖͘ͳΔ

  91. 91 Ճ๏త ৐๏త قઅੑͷେ͖͞͸Ұఆ قઅੑ͸ݩͷ஋ʢτϨϯυʣʹൺྫͯ͠େ͖͘ͳΔ

  92. قઅੑͷϞʔυʢՃ๏తɺ৐๏తʣͷ੾Γସ͑ 92 ΞφϦςΟΫεɾϓϩύςΟͷقઅੑͷ ϞʔυͰɺՃ๏త͔৐๏త͔Λ੾Γ͔͑ Δ͜ͱ͕Ͱ͖Δɻ

  93. 93 Ճ๏త ৐๏త

  94. 94 Ճ๏త ৐๏త ৐๏తͳقઅੑΛ࢖ͬͨ༧ଌͷ΄͏ ͕࣮ଌ஋ʹ༧ଌ஋͕௥ਵ͍ͯ͠Δɻ

  95. 95 Ճ๏త ৐๏త ςετ݁ՌΛΈΔͱɺ৐๏తͳقઅੑΛ࢖ͬͨ༧ଌͷ΄͏͕ɺશͯͷࢦඪ ͰɺΑ͍݁Ռʢখ͍͞஋ʣʹͳ͍ͬͯΔɻ

  96. قઅੑͷϞʔυʢՃ๏తɺ৐๏తʣͷൺֱ 96 قઅੑͷύλʔϯ͕ݟ΍͍͢Α͏ʹɺ࣌ ؒͷ୯ҐΛ͍ͬͨΜWEEK(ि)ʹͯ͠ɺ قઅੑͷϞʔυΛՃ๏తͱ৐๏తͷؒͰ ੾Γସ͑ͯɺ༧ଌ݁ՌΛൺֱͯ͠ΈΔɻ

  97. ݄୯Ґʹ໭ͯ͠ɺςετ݁ՌΛൺֱ 97 ࣌ؒͷ୯ҐΛݩͷMON(݄)ʹͯ͠ɺقઅ ੑͷϞʔυΛՃ๏తͱ৐๏తͷؒͰ੾Γ ସ͑ͯɺςετ݁ՌΛൺֱͯ͠ΈΔɻ

  98. 98 ʮ܁Γฦ͠ʯΛ࢖ͬͯෳ਺ͷϞσϧΛ࡞Γɺ ͦΕΒΛൺֱ͢Δ

  99. 99 Ϛʔέοτ͝ͱʹ༧ଌϞσϧΛ࡞Γ͍ͨͷͰɺʮ܁Γฦ͠ʯʹ ’Market’ ྻΛબͿ

  100. 100

  101. • RMSE, MAEͰݟΔͱɺAfrica (ΞϑϦΧ)ͷ΄͏͕Asia Pacific (ΞδΞଠฏ༸஍ Ҭ)ΑΓ༧ଌ஋ͱ࣮ଌ஋ͷ͕ࠩখ͍͜͞ͱ͕෼͔Δɻ • ͜ΕΛ΋ͬͯɺAfricaͷํ͕Α͘༧ଌͰ͖͍ͯΔͱݴ͑ΔͩΖ͏͔ʁ Ϟσϧͷ༧ଌਫ਼౓ͷࢦඪΛൺ΂Δ

    101
  102. • Asia Pacific (ΞδΞଠฏ༸஍Ҭ) ͷํ͕Africa (ΞϑϦΧ)ΑΓച্ֹ͕େ͖͍ͷ ͰɺRMSE, MAE͕େ͖͘ͳΔͷ͸౰ͨΓલͱݴ͑Δɻ • ஋ͷεέʔϧʹ͕ࠩ͋Δͱ͖͸ɺRMSEɺMAEʹΑΔ༧ଌੑೳͷൺֱ͸ҙຯΛ

    ͳ͞ͳ͍ɻ Ϟσϧͷ༧ଌੑೳͷαϚϦͷදࣔ 102
  103. • MAPE͸ɺ༧ଌ஋ͱ࣮ଌ஋ͷࠩΛɺ࣮ଌ஋ͷύʔηϯτͰදͨ͠΋ͷɻ • ࣮ଌ஋ͷ΋ͱ΋ͱͷେ͖͞ʹؔ܎ͳ͘༧ଌੑೳͷൺֱ͕Ͱ͖Δɻ • RMSE, MAEͷେ͖͔ͬͨAsia Pacificͷ΄͏͕ɺMAPE͸ΑΓখ͘͞ɺ࣮͸ AfricaΑΓΑ͍༧ଌ͕Ͱ͖͍ͯͨ͜ͱ͕෼͔Δɻ Ϟσϧͷ༧ଌੑೳͷαϚϦͷදࣔ

    103
  104. • MAPE΋ಉ͡Α͏ʹɺ࣮ଌ஋ͷ΋ͱ΋ͱͷେ͖͞ʹؔ܎ͳ͘༧ଌੑೳͷൺֱ͕Ͱ͖Δɻ • MAPE͸ɺ஋͕0Λ·͍ͨͩΓɺ0ʹۙ͘ͳΔͱ͖͸ෆ҆ఆʹͳΔ͕ɺMASEʹ͸͜ͷ໰୊͸ͳ͍ɻ • MASEͰݟͯ΋ɺRMSE, MAEͷେ͖͔ͬͨAsia Pacificͷ΄͏͕ɺ࣮͸ΑΓΑ͍༧ଌ͕Ͱ͖͍ͯͨ͜ͱ ͕෼͔Δɻ Ϟσϧͷ༧ଌੑೳͷαϚϦͷදࣔ

    104
  105. ֎෦༧ଌม਺

  106. ച্ʢSalesʣͱ૬ؔؔ܎ʹ͋Δม਺͕͋ͬͨͱͯ͠ɺ͞Βʹͦͷ ม਺ͷকདྷͷ஋Λ༧ଌͰ͖Δɺ·ͨ͸ίϯτϩʔϧͰ͖Δͱͨ͠ Βɺച্ʢSalesʣΛ͞Βʹྑ͍ਫ਼౓Ͱ༧ଌͰ͖ΔͷͰ͸ͳ͍͔ʁ

  107. Sales ͱ Sales Comp. ͸૬͍ؔͯ͠ΔΑ͏ͩɻ

  108. Sales ͱ Marketing ΋૬͍ؔͯ͠ΔΑ͏ͩɻ

  109. Sales ͱ Discount (Avg) ͸͋·Γ૬ؔͯͦ͠͏ʹͳ͍ɻ

  110. ϚʔέςΟϯά ࣍ͷ3ϲ݄ɺͲΕ͚ͩ޿ࠂʹ͓ۚΛඅ΍͔͢ίϯτϩʔϧͰ͖Δͱͨ͠ Βɺͦͷ৘ใΛϞσϧʹ૊ΈࠐΉ͜ͱ͸Ͱ͖ͳ͍͔ʁ ྫ

  111. ఱؾ ച্͸ఱؾʹΑͬͯӨڹΛड͚Δͱ͢Δͱɺ࣍ͷ10೔ؒͷؾԹ΍߱ਫ֬཰ Λ༧ଌͯ͠ɺͦΕΒΛച্ͷ༧ଌϞσϧʹ૊ΈࠐΉͱɺ͞Βʹྑ͍ਫ਼౓Ͱ ച্͛Λ༧ଌͰ͖ΔͷͰ͸ͳ͍ͩΖ͏͔ɻ ྫ

  112. ม਺ʢྻʣΛ֎෦༧ଌม਺ͱׂͯ͠౰ͯɺ༧ଌϞσϧΛ ࡞Δ͜ͱ͕Ͱ͖Δɻ Prophet͸֎෦༧ଌม਺ʢྫɿϚʔέςΟϯάඅ༻ʣ͕ λʔήοτม਺ʢྫɿച্ʣΛ༧ଌ͢ΔͨΊʹ໾ཱ͔ͭ Λௐ΂ɺͦͷ֎෦༧ଌม਺ͷ܎਺Λ୳͠ग़͢ɻ ֎෦༧ଌม਺

  113. τϨϯυͱ೥पظͷقઅੑΛ΋ͱʹͨ͠༧ଌϞσϧ

  114. ϕʔεϞσϧͷධՁ

  115. τϨϯυͱ೥पظͷقઅੑͱച্ใुΛ΋ͱʹͨ͠༧ଌϞσϧ

  116. ϞσϧͷධՁ ϕʔεϞσϧ ϕʔεϞσϧʹച্ใुΛ෇͚଍ͨ͠

  117. τϨϯυͱ೥पظͷقઅੑͱϚʔέςΟϯάඅ༻Λ΋ͱʹͨ͠༧ଌϞσϧ

  118. ϕʔεϞσϧʹച্ใुΛ෇͚଍ͨ͠ ϕʔεϞσϧʹϚʔέςΟϯάඅ༻Λ෇͚଍ͨ͠

  119. ࣍ͷ3ͭશ෦଍ͯ͠ΈΔ Sales Comp., Marketing, and Discount

  120. With Sales Comp., Marketing, Discount

  121. The forecasting model quality has improved for a little bit.

    ϕʔεϞσϧʹച্ใुΛ෇͚଍ͨ͠ ϕʔεϞσϧʹച্ใुɺϚʔέςΟϯάඅ༻ɺׂҾ཰Λ෇͚଍ͨ͠
  122. ޮՌλϒͷԼͰ͸ɺͦΕͧΕͷقઅੑͱ༧ଌม਺͕༧ଌ஋ʹͲͷΑ͏ʹ ӨڹΛ༩͑Δͷ͔ΛݟΔ͜ͱ͕Ͱ͖Δɻ

  123. ม਺ॏཁ౓λϒͷԼͰ͸ɺ࢖ΘΕͨقઅੑͱ༧ଌม਺ͷ͏ͪͲͷม਺͕ ΑΓӨڹ͕ڧ͍ͷ͔ΛݟΔ͜ͱ͕Ͱ͖Δɻ

  124. Q & A

  125. ࣍ճηϛφʔ

  126. σʔλαΠΤϯε X 3/5 (໦) 1PM (೔ຊ࣌ؒʣ

  127. None
  128. 5݄։࠵ܾఆʂ ฏ೔൛ɿ26೔ʢՐʣɺ27೔ʢਫʣɺ28೔ʢ໦ʣ ि຤൛ɿ23೔ʢ౔ʣɺ24೔ʢ೔ʣɺ30೔ʢ౔ʣ

  129. • ϓϩάϥϛϯάͳ͠ RݴޠͷUIͰ͋ΔExploratoryΛ෼ੳπʔϧͱͯ͠࢖༻͢ΔͨΊडߨத͸ɺϏδωεͷ ໰୊Λղܾ͢ΔͨΊʹඞཁͳσʔλαΠΤϯεͷख๏ͷशಘʹ100ˋूதͰ͖Δ • ෼ੳπʔϧͷϕϯμʔϩοΫΠϯͳ͠ ExploratoryͰͷ࡞ۀ͸શͯಠཱͨ͠ΦʔϓϯιʔεͷR؀ڥͰ࠶ݱ͕Մೳ • ϏδωεͰ࢖͑ΔࢥߟྗͱεΩϧͷशಘ σʔλαΠΤϯεͷεΩϧशಘ͚ͩͰͳ͘ɺσʔλ෼ੳʹඞཁͳࢥߟྗ΋शಘͰ͖Δ

    ಛ௃
  130. ࿈བྷઌ ϝʔϧ kan@exploratory.io ΢ΣϒαΠτ https://ja.exploratory.io ϒʔτΩϟϯϓɾτϨʔχϯά https://ja.exploratory.io/training-jp Twitter @KanAugust