A player’s value can be boiled down to two functions. First is marginalization, i.e. the relationship between spending patterns and profit. How much revenue can I expect given a player’s spending patterns? What and how much cost should be attributed to such patterns? Second, the prediction, i.e. the forecast of a player’s spending patterns. What spending patterns can I expect from a player’s future visit? How many future visits can I expect? What factors influence spending patterns and visit frequency?
Customer valuation affects many parts of the organization. For the revenue manager, gaming value is the primary input to yield hotels rooms in a casino. Players are segmented based on their value, and the revenue managers forecast demand and recommend rates/COMPs at the segment level. Thus, an imprecise valuation methodology, e.g. one that overvalues players by 12%, will directly affect their pricing decision. The customer value also impacts marketing decisions, e.g. promotional offers, player development as well as financial ones since the value of the player database is a financial asset.
Customer valuation for a casino is particular for two reasons. Randomness make marginalization harder. Players mostly control how much and how they wager while randomness mostly drives how much they win. The cost side is driven by the redemption of COMPs and free play. As a result, identical wagering patterns can result in highly variable profit margins.
On the other hand, great, vast data makes prediction easier. Spending patterns of each visit are dutifully recorded and stored. A significant proportion of players are repeat or frequent customers and historical patterns are a good indicator of future visits and spend.
The win measures the actual gaming revenue. It lines up with the gaming revenue listed in the financial reports. In a sense, it measures the revenue from what was wagered by the player. The win is an unfair metric: when using the average win over the last 5 trips to forecast the value, 20% of the players had a negative value (18K out of 90K) and 69% of them had a positive win on the next trip (13K out of 18K). Plus its high volatility makes it particularly unsuited for forecasting. The theoretical win estimates the expected gaming revenue. It estimates the revenue from what was wagered by the player if we removed all randomness from the games
The theoretical win is usually underestimated. Players with limited bankroll can get unlucky and lose it all before having the opportunity to generate a significant amount of theoretical win. The worth refers to theoretical win once it has been adjusted. It is commonly defined on a trip basis as the maximum of the theoretical win -- 40% of win, 40% being the worth coefficient. The net worth is an estimation of gaming profit and accounts for taxes, free play and COMP dollars.
Most players are repeat customers in frequency markets. Their past spending behavior is a good indicator of future one. The worth for a future trip is generally forecasted using the Average Daily Worth (ADW) over the last 12 months: Value = sum [Net worth over 12M] / sum [gaming days over 12M]
How to Assess a Customer Valuation Methodology
How can a CV methodology be objectively assessed? As far as the marginalization is concerned, one should make sure that the worth aligns with the win when summed across a large sample of trips. Free play usually needs to be subtracted from theoretical win to align with the actual one, so it needs to be considered or both worth and value may be overestimated. As far as the worth formula is concerned, using a blanket 40% in the worth adjustment often overshoots and leads to a worth that is much higher than the win. Several worth coefficients should be tested offline until an adequate one is found.
The prediction should be assessed based on its accuracy, meaning the forecast error should be minimized. There are multiple ways to define accuracy. If a a player value was forecasted to be $257 but the player was actually worth $123, the error is -$134. The percentage error is -52%. The absolute percentage error (APE) is 52%. The APE is weighted and averaged across all trips. The resulting metric –WAPE – can be used to measure the forecast accuracy of the CV.
Let’s consider the financial impact of forecast accuracy. A hotel guest whose value is overestimated may displace higher value players or be eligible for unearned hotel rates. In a case study, it was simulated that a decrease of 3% in WAPE for the entire casino floor resulted in an annualized $3M (14%) reduction in over-forecasting for hotel guests.
One way to improve forecast accuracy is to run simulations. For instance, a sample of 16,500 players is selected randomly from a regional hotel casino in the U.S. These players made 45,799 trips over the last 3 years. 16,057 of these trips occurred in the last year and were used as a hold-out sample. The valuation was simulated using historical trips preceding each of the held-out one.
The resulting WAPE is 73.77% which seems high but keep in mind that player worth is highly variable both trip over trip and across players. Besides a significant portion of players are infrequent and thus have very little history. The value for first time players is assumed to be $0, which further degrades the accuracy.
More interesting were the improvements measured when parameters of the valuation were changed. Assuming that the worth was proportional to the length of the trip resulted in a better accuracy than assuming that it was independent (73.77% vs. 82.93%). As long as the formulas are linear combinations, forecasting slot and table separately has no impact. The same applies to forecasting profit directly vs. forecasting revenue and cost separately.
Using the date of the last trip as point of origin rather than the booking date improved the WAPE (71.16% vs. 73.77%). This could be because it ensures that we always consider at least one historical trip for infrequent players. When changing the historical horizon from 12 months to 6, 18 and 24 months, 18 months was the most accurate (70.93% vs. 72.33%, 71.16%, 70.93% and 71.07% for 6, 12 and 24 months respectively).
Limiting historical observations using a number of trips rather than time span improved the results, and using the last 10 trips was more accurate than using the last 5 (70.61% for last 10 vs. 69.80% for last 5). Finally, reducing the weight of each trip exponentially further reduced the WAPE (68.92% vs. 69.80%)
These results are not necessarily the same for all properties and group of players. Thus it is recommended that each property runs its own simulations to determine what results in the most accurate valuation. Potentially, players can be segmented and the forecast fine-tuned at the segment level.