Lessons from two case-studies: How to Build Accurate and models

Transcription

1 Lessons from two case-studies: How to Build Accurate and models Palisade Risk Conference New Orleans, LA Nov, 2014 Huybert Groenendaal, PhD, MBA Managing Partner EpiX Analytics

2 Two objectives Quick overview stories: 1. Forming the right partnership agreements 2. Picking an optimal project portfolio Share several technical lessons

3 Story #1.the question Call from a large multinational company Question: We currently sell stuff, but we d like to understand how we can potentially make more money.. help? Yesterday!

4 Story #1. the background Successful company with specific technologies and expertise Sell products B2B that require upfront R&D Margin business How about taking more risk, and sharing in some of the upside?

5 Story #1. design the model First make sure to understand all risks and (critically important) all relationships! Then develop blue-print of the model Perceived accuracy Complexity Time & effort Accuracy Transparency 0 0 Size of the model / amount of detail

6 Story #1. Build the model Build P&L model Variety of risks and uncertainties: R&D costs; R&D timeline; Product pricing; Market share/unit sales; Competitive situation at product launch; Note: Risk and uncertainties can both be up and downsides

7 Story #1. what about correlations? Deserves great effort! But typically. no rank-order correlations are used! Huh? Important rule in Monte Carlo simulation: Every iteration (all 10,000!) has to be a possible future scenario.

8 Why should we all value our relationships? Omitting to include relationships typically results in the under-estimation of risk; user only use the rank-order correlation, but there are many alternative methods to include relationships: Envelop method M0146 Lookup tables; M0275 Logical relationships (IF-statements etc.) M0097 The bootstrap M0444 Relationship based on a regression (linear or non-linear) General recommendation: Only use rank-order correlation if no other correlation methods can be used!

9 ModelAssist Free Monte Carlo training and reference tool: Risk Analysis Training tool Based on 20+ years of risk analysis consulting and training expertise Within ModelAssist, page numbers are Mxxxx. For example, M0407 will get you to the Selecting the appropriate distributions for your model

10 Some example relationships Example models are in ModelAssist Conditional logic Envelop method Regression Bootstrap Lookup tables

11 Story #1.evaluating partnerships? Different partnership structures and terms were they overlaid on top of the overall risked P&L of the project Example: Partnership 1 Partnership 2 Partnership 3 Partnership 4 R&D cost-sharing Cost 1 50% 100% 100% 100% Cost 2 0% 0% 50% 100% Margin on product 200% 50% 0% -100% Profit sharing % 0% 10% 25% 50%

12 Value (US$) Story #1. the result? $35,000,000 NPV Agreement 1 NPV Agreement 2 NPV Agreement 3 NPV Agreement 4 NPV Agreement 5 $30,000,000 $25,000,000 $20,000,000 $15,000,000 $10,000,000 $5,000,000 $- $(5,000,000) $(10,000,000) Model provides many more decision-supporting results and graphs. Lets have a look: The model

13 The moral of case-study can be a powerful decision-supporting tool to understand the risk-reward picture of alternative partnership agreements: Product / business development Private equity / venture capital Joint ventures, etc. Model blue print: Understand uncertainties before building models Don t over-complicate models Many ways of including relationships within a model Rank-order correlations often last resort

14 Story #2.the question Call from a large multinational company Question: We have a large stochastic optimization model, but it only gives us an answer after waiting for 3 days. and the answer isn t even that optimal. Any thoughts / suggestions / ideas?

15 Story #2. the background Large oil and gas company Drilling of wells with large budget but: Still more wells than budget allowed; Not enough drilling equipment FTE s limited Some wells could only be accessed in winter Company had built a large optimization model that included uncertainty (i.e. Monte Carlo simulation)

16 Story #2. the optimization model Conceptually similar to the following model: The model

17 Continue for example 1000 trials or 5 min... How does stochastic optimization work? What does Risk Optimizer do? 1. Run full Monte Carlo simulation of the model 2. Look if constraints have been met (sometimes done before step 1) 3. Check if objective found is superior to what was found before: If so, save it! 4. Select another set of adjustable ranges/decision variables

18 Story #2. Real-life optimization models: Project portfolio with 25 project, how many possible portfolios are there? How about 50 projects? How about 200 projects?

19 Story #2. So, what do we do? Speed up MC simulations: Take less time evaluating each place Constrict the decision-space: Don t spoil time looking in eastern Colorado Discretize solution-space Don t look at every square foot, but survey per acre

20 Speeding up evaluation time: Easy wins such as: Using faster functions within Excel; Running fewer iterations; Less/no formatting; Minimizing the model to only necessary calculations More advanced options: Splitting the simulation from the optimization Use probability calculations or Central Limit Theorem

21 Constricting the decision-space: Adding additional constraints For example, in addition to a budget ceiling, add a budget floor The model Related to this, it often can help speed to start at a close to optimum solution

22 The moral of case-study #2. Stochastic optimization can help management uncover more value within business constraints: R&D portfolios; Drilling portfolios; Blending problems, etc. However: Optimization models can quickly become slow; Care is needed in building models that provide robust answers in acceptable time

23 The moral of both can provide powerful insight to help with: 1. Forming the right partnership agreements 2. Picking an optimal project portfolio Some technical lessons: 1. Importance of blue-prints 2. Relationships models 3. Techniques to building faster optimization models 4. Don t run away from mountain lions.!

24 Thank you! Dr. H.Groenendaal Managing partner EpiX Analytics LLC P: