Latest / The Yield Doctor with James Deaker / Inventory Forecasting Made Easy: 5 Rules for Digital Publishers

Transcript

Inventory Forecasting Made Easy: 5 Rules for Digital Publishers · The Yield Doctor with James Deaker

• 0:00
  If you ask an AI agent right now to give
• 0:03
  you an inventory forecast for next
• 0:06
  quarter, it will probably deliver a
• 0:08
  beautiful chart to you with that
• 0:10
  inventory forecast in about 3 seconds
• 0:13
  flat. But if that forecast is built on
• 0:17
  the wrong business assumptions, it is
• 0:19
  just a high-speed way to scramble your
• 0:21
  pacing, to blow up your quarterly
• 0:24
  financial plan, or potentially to leave
• 0:27
  millions of dollars of revenue on the
• 0:29
  table. Getting a forecast is not the
• 0:33
  hard part anymore. Knowing what forecast
• 0:37
  you actually need is where the game is
• 0:39
  won and lost. And so what I'm going to
• 0:41
  be talking about today is what you as
• 0:44
  the human operator, as the human in the
• 0:47
  loop, need to ask of your agents at the
• 0:51
  start of the process and what you need
• 0:53
  to be looking for in your output. And
• 0:56
  these concepts are essential whether you
• 1:00
  are in ad operations, inventory
• 1:02
  management, yield or even finance.
• 1:05
  anybody who is responsible for
• 1:08
  understanding or producing inventory
• 1:10
  forecasts.
• 1:12
  By the end of this video, you will have
• 1:14
  five specific concepts that you need to
• 1:16
  be thinking about when you work with
• 1:19
  your algorithm or your AI agent or the
• 1:22
  platform around inventory forecasting as
• 1:25
  a publisher of digital advertising.
• 1:28
  Before I begin, there's one important
• 1:30
  consideration that a lot of
• 1:32
  organizations skip right past, which is
• 1:35
  why are you even doing the supply
• 1:38
  forecast in the first place? The most
• 1:41
  immediate reason for a good supply
• 1:44
  forecast is to ensure that you have the
• 1:47
  inventory available to deliver on the
• 1:50
  commitments you have made for your
• 1:53
  managed advertising campaign. But just
• 1:56
  as important are considerations like
• 1:58
  ensuring that you know what supply you
• 2:01
  will have over a certain period so that
• 2:04
  your ad server paces correctly or that
• 2:07
  you have enough supply available to meet
• 2:09
  your financial plan for the quarter or
• 2:12
  for the year. And what I just ran
• 2:14
  through are a number of very different
• 2:16
  business questions. For example, a
• 2:19
  forecast can be used to check whether a
• 2:22
  guaranteed campaign will deliver and
• 2:25
  that may be much more granular than a
• 2:28
  forecast that is used for quarterly
• 2:30
  financial planning purposes. So the
• 2:33
  first place to start is not with the
• 2:35
  model. It's about thinking about what is
• 2:38
  the decision that you are trying to
• 2:40
  support.
• 2:42
  Regardless of the use case, the first
• 2:44
  place to start thinking about that is
• 2:46
  the time periods that are important for
• 2:48
  your business because this informs the
• 2:51
  level of detail that you need in your
• 2:53
  forecast.
• 2:55
  For example, you may know that your
• 2:57
  audience peaks in the early morning and
• 3:00
  midafter afternoon in terms of views and
• 3:02
  the traffic at this time is
• 3:04
  significantly greater than say the early
• 3:06
  hours of the morning. And if you're a
• 3:09
  publisher who is selling specific hourly
• 3:12
  slots, this is important. Maybe you have
• 3:15
  a lot of fast food advertisers as
• 3:18
  clients and you are selling impressions
• 3:20
  around the lunch hour rush. However, for
• 3:25
  the vast majority of digital publishers,
• 3:28
  this type of information is absolutely
• 3:31
  meaningless. It's not going to change
• 3:33
  how your sales team sells. It's not
• 3:35
  going to change how your finance team
• 3:37
  plans the quarter and it's probably not
• 3:39
  going to change how the yield team
• 3:41
  manages the business. And the lesson in
• 3:44
  that is that more detail is not always
• 3:48
  better. What you care about in most
• 3:51
  cases is the total number of impressions
• 3:54
  of a certain type that are available on
• 3:57
  a daily basis. Now, I threw a lot of
• 4:01
  information into that phrase of a
• 4:02
  certain type. And what I mean is what is
• 4:06
  the level of aggregation that your sales
• 4:09
  team sells at or which you're making
• 4:11
  available programmatically. And usually
• 4:14
  it's going to correspond to the type of
• 4:16
  products you represent on your rate
• 4:18
  card. At a bare minimum, you need to
• 4:21
  have a different forecast for each type
• 4:24
  of format for display ads, for native
• 4:27
  ads, for video ads, and for each
• 4:29
  individual ad size. Within that, you
• 4:33
  also need to have a different inventory
• 4:36
  forecast for each subsite or sub app
• 4:39
  that you support.
• 4:42
  For example, I assure you that Yahoo,
• 4:44
  the website, would have different
• 4:46
  inventory forecasts for Yahoo Sports,
• 4:50
  Yahoo News, Yahoo Fantasy Football, and
• 4:54
  so on. And this is where the human
• 4:56
  operator matters. An I an AI system can
• 5:01
  generate a forecast at almost any level
• 5:04
  of detail, but the human has to decide
• 5:07
  the level of detail that actually
• 5:09
  matches how the business operates.
• 5:13
  So, I've told you that outside of
• 5:16
  perhaps ad server campaign pacing for
• 5:19
  most business processes, you don't need
• 5:22
  to worry about forecasting at the hour,
• 5:24
  minute, and second. But clearly you do
• 5:27
  need to understand at the day, week and
• 5:30
  month. And the key is understanding the
• 5:33
  unique dynamics in your audience. Does
• 5:36
  your audience show up more during the
• 5:38
  week or the weekend?
• 5:41
  Well, this will be a function of the
• 5:43
  type of content that you show. Sports
• 5:45
  sites tend to have higher audience
• 5:48
  counts on the weekend. News and finance
• 5:51
  sites tend to be higher during the week.
• 5:54
  Another question to ask yourself, does
• 5:56
  your audience have an annual cycle? New
• 5:59
  sites tend to be fairly constant
• 6:01
  throughout the year, whereas something
• 6:03
  like fantasy football is going to have a
• 6:05
  huge uplift during the NFL season and
• 6:08
  particularly in the first few weeks and
• 6:10
  be dead for much of the rest of the
• 6:12
  year. And these are the types of
• 6:15
  dynamics where a human needs to ensure
• 6:17
  that the algorithm is picking up the
• 6:19
  right signals. The second main area
• 6:22
  where a human should be helping an
• 6:23
  algorithm or an AI agent with
• 6:26
  forecasting is on long-term trends as
• 6:29
  they pertain to the company. So, for
• 6:32
  example, if you're a publisher who had
• 6:34
  search traffic in a search bar on your
• 6:36
  website, for many years, search traffic
• 6:40
  was steadily increasing year on year.
• 6:43
  Unfortunately, now much of the traffic
• 6:47
  has moved away from search bars on
• 6:49
  publisher sites and is moving
• 6:52
  increasingly into direct agentic
• 6:55
  searches. And that means that search
• 6:57
  traffic is predictably decreasing
• 7:01
  instead of increasing across many
• 7:03
  publishers. And a good analyst needs to
• 7:06
  be on top of these types of dynamics to
• 7:08
  understand how to update future
• 7:11
  forecasts.
• 7:13
  This is one of the areas where
• 7:15
  historical data can mislead you. The
• 7:18
  model may see what happened last year
• 7:20
  and project it forward. But the human
• 7:23
  has to understand the underlying
• 7:26
  business and understand that it has
• 7:28
  changed. Maybe search traffic is
• 7:32
  declining. Maybe a product is being
• 7:34
  redesigned. Maybe a major distribution
• 7:37
  partner is going away. Maybe a new app
• 7:41
  release is expecting to change user
• 7:44
  behavior.
• 7:46
  These are not just technical forecasting
• 7:49
  issues. These are business judgment
• 7:51
  issues and they are these are also areas
• 7:55
  where I see some teams dedicating too
• 7:59
  much time or in another way of putting
• 8:00
  that is wasting too much time. One area
• 8:04
  is around trying to forecast spikes
• 8:06
  around specific events. And this might
• 8:09
  be an example where a story
• 8:12
  unpredictably suddenly goes viral and
• 8:15
  generates a huge volume of traffic. In
• 8:18
  this type of situation, the head of
• 8:20
  sales or revenue realize that a lot of
• 8:23
  these impressions aren't being fully
• 8:25
  monetized and starts yelling at the
• 8:28
  operations team to give better
• 8:29
  forecasts.
• 8:31
  But in most cases, there isn't actually
• 8:34
  a lot of value in spending time on
• 8:37
  problems like this because of the
• 8:39
  unpredictable nature and the total
• 8:41
  volume of impressions that can be
• 8:43
  monetized. Time and attention is usually
• 8:46
  spent better spent elsewhere. An
• 8:49
  exception to this may be if the spike
• 8:52
  occurs at a very specific and
• 8:55
  predictable time from one year to the
• 8:57
  next and you have historical logs that
• 9:00
  allow you to predict this. Think of
• 9:02
  things like Black Friday, Cyber Monday,
• 9:05
  the Super Bowl and so on. However, even
• 9:09
  in these cases, I would question whether
• 9:11
  the amount of time and effort that many
• 9:13
  teams put into this is warranted for
• 9:16
  most publishers because the question
• 9:18
  should always be, will this forecast
• 9:21
  change a decision within the company? If
• 9:24
  the answer is no, you may be creating
• 9:27
  complexity without creating value.
• 9:30
  I've create covered a number of
• 9:32
  different concepts here. So, let me
• 9:34
  summarize.
• 9:35
  Inventory supply forecasting is hard
• 9:38
  within digital media and so don't make
• 9:41
  it harder than it needs to be. First,
• 9:44
  you need to understand why you are
• 9:46
  forecasting in the first place. Are you
• 9:49
  doing it to manage campaign delivery,
• 9:51
  help the ad server pace correctly,
• 9:54
  support the sales team, or to validate
• 9:56
  the financial plan?
• 9:58
  Secondly, make sure that you understand
• 10:01
  the time period that is relevant for
• 10:03
  your business and your advertisers.
• 10:06
  Third, make sure that you understand the
• 10:08
  level of aggregation that is
• 10:10
  appropriate, meaning what product types,
• 10:13
  ad formats, ad sizes, site levels, and
• 10:16
  app levels actually matter. Fourth, help
• 10:20
  the algorithms by incorporating the
• 10:23
  right long-term business trends. And
• 10:26
  fifth, and finally, don't waste time
• 10:29
  forecasting unpredictable spikes unless
• 10:32
  you have to. The biggest takeaway is
• 10:35
  this. Don't ask your AI agent just for a
• 10:39
  forecast. Ask for a forecast that
• 10:43
  matches the business decision that you
• 10:46
  need to make. I'm James Deika. I am the
• 10:49
  yield doctor. Subscribe to join me next
• 10:51
  time.