 Hello everyone, thank you for being here. I will address today the concept of systematic fundamental arbitrage in equity. I will answer all the questions as mentioned at the end of the presentation, quantitative or qualitative. What I mean by equity fundamental arbitrage is alpha generation resulting from inconsistencies or imbalances between the market value and the equity value of a company. We can measure these imbalances and inconsistencies in reference to a static number or in reference to comparable companies. So if we have a PE ratio, price to earnings ratio above a given threshold, we would deem the stock overvalued. But if one firm's PE ratio is above another firm's PE, then the first stock is deemed more expensive than the second. What we call relative value analysis here is exactly that. It's a peer to peer analysis. Now when I started to work on a similar project 10 years ago at Société Générale, I had just graduated from college and I had a real obsession with applied mathematics and financial models in general. And I took the components of the S&P 500, I was using fax set at the time, I'm not promoting fax set here, but I had easily over 10 years history and I thought now that I had every input I could hope for, everything I need to do is concentrate on theory. And I was of course very wrong and whatever the project data control is 80% of the job. Now equity fundamental arbitrage cannot be conceived without a thorough understanding of both the disclosure process and the nature of the firm's financial statements. Any meaningful approach in equity is a pragmatic approach. We need to understand how material information is disclosed, when it is supposed to be disclosed and how to read it. So if the rules of disclosures are different, data will most likely be different. Equity under IFRS is defined differently from equity under US GAAP. So data integrity has a legal and regulatory aspect here that I explore through the accounting treatment of defined benefit pension plans. Once all disclosure rules and market rules are accounted for, I can factor in my own rules to eventually design a systematic strategy. So I'll go through the specifics of the residual income model as presented by Gebart in 2001 and then I will describe the calibration performed by Fama in French. I will discuss a simpler model and a different approach for its calibration, one that is adapted for a relative value analysis. Finally, I will test the robustness of our scoreings, discuss briefly the challenges of a smart beta allocation and in the end challenge the results. So throughout this presentation, I will complete the strategies profile, its objectives and constraints. You can see here, we have the two standard objectives of return and risk and four investment constraints, time horizon, liquidity constraints, legal and regulatory and unique circumstances that are really specific to equity shares. So far, I have my return objective of consistent alpha generation through statistical analysis of financial statements and data in general. So let's fill in some more of the investment constraints first. Increasing stocks requires rigor, especially if you are engaging in long-short or market useful strategies. Execution is an essential part of the process. Well, good news first, brokerage fees and transaction fees in general have dropped by half these last few years due mostly to an increasing competition among brokers and prime brokers. But anyone who has traded stocks or stock derivatives knows that trend following strategies can suffer high implicit costs, high slippage costs. A trader can spend hours chasing a price and in the end these costs might offset any expected benefit. So designing liquidity provider strategies rather than trend following strategies is always a better option when dealing with stocks. And by liquidity provider strategies, I mean mini reverting strategies. We also know that shorting a stock close to a corporate action right after an IPO can be highly expensive, not to say sometimes impossible. Borrowed costs can persist at high levels, sometimes more than 50%, months after an IPO. The recent Groupon or Facebook IPOs are proof of that and there was an extraordinary high demand for these stocks. But in current markets and in standard market conditions more than 90% of the total return swaps imply borrow costs significantly below 2%. We also have legal constraints generating market imperfections such as the optic rule on the Japanese markets or the short selling restrictions on financial stocks in the midst of the crisis in 2008. Finally, it's important to consider account margin requirements for an optimized cash management. At best and here because we are not dealing with event driven strategies we should avoid critical firms and low level stocks that are more exposed to potential short squeeze. Now all these market specifics make investors reluctant in operating with equity shares. They are among the reasons why strategies and indices defined as fundamental arbitrage have mostly failed. Now for liquidity constraints I can fill in here my constraints, I'm better off looking at the float or the trading volume rather than the market capitalization and I am looking for liquidity provider strategy rather than a mean reverting strategy. For unique circumstances we should pay attention to borrow costs and account margin requirements, we should also avoid critical firms and speculative stocks. Now note that I entitled the first subsection a control disorder for all the reasons that I have mentioned so far and for more reasons that I will mention later. I will try to bring some measure of control to this disorder and eventually lay out the steps towards a systematic investment solution. Now apart from market considerations a large portion of this disorder is explained by the release of material information on earnings announcement dates. Some other portion is explained by unscheduled release of material information such as profit warnings and results guidance. Now before playing the game and it's a large scale game over thousands of firms we absolutely need to go through the rules. We are working on the design of a systematic strategy and a systematic strategy is based on the rules, the existing rules and your own compatible rules. Now most existing rules are obviously set up by the SEC for the US market and by the relevant authorities in Europe. Now the first simple rule defines three categories of filers depending on their public foot. A firm with a public foot above $700 million will file the annual 10K within the deadline of 60 days after the end of the fiscal year and the quarterly 10Q reports within 40 days after the end of the quarter. These firms are referred to as a large accelerated filers. Now for accelerated filers with a public foot below $700 million and above $75 million, the firm must file within 75 days for the 10K annual filings and within 40 days from the end of the quarter period. Now obviously the SEC applies larger deadlines for lower public foot because the filing deadline is a constraint and like any constraint it comes at a cost. A fixed cost that doesn't seem to be justified by the benefits that investors would obtain from earlier access to the reports. I invite you all to read the December 2005 revisions to the accelerated filer definition and accelerated deadlines for filing periodic reports. You'll find there a complete cost-benefit analysis. Now apart from the 10K and 10Q periodic reports we have current 8K filings that are more generally produced to disseminate material information to the public in a timely manner. Now a firm must file the 8K within four business days of the event's occurrence. You can refer to the general instructions included in any 8K or you can refer to the Code of Federal Regulations CFR for all details on reporting requirements. The Code of Federal Regulations lists nine sections that cover 31 items and each item is a reportable event. Now among these items we have a particular interest in item 2.02 results of operations and financial condition and item 9.01 financial statements and exhibits. Most of the accounting data we need is already disclosed there within a few business days of the announcement. Now it is completed by the full quarterly and annual reports available within two months from the end of the quarter period or the end of the fiscal year. Now gathering what I have mentioned so far this is how the schedule looks like. It's a simple schedule. I have the end of period, I have my 8K filed at the announcement date, I have the quarterly and annual 10Q and 10K filings at deadlines depending on the firm's float. Now even though data included in the 8K is sometimes incomplete or limited it can be completed with information disclosed at the conference call or on the firm's website. It can also be in the most extreme cases extrapolated or estimated. I'm thinking of equity and cash flow from operations and so on. Now why is this so important? We are looking to managing a portfolio of stocks selected among thousands of companies. I would like a practical way for detecting when my data for a company is outdated or when this data is due so I can actually optimize the decision making process. I can automate all I can to concentrate my efforts on risk management and execution. So far if we leave it this way it's not too bad. To spice it up a little bit there is always in every game at least every interesting game a tricky section to the rules and this section here is about restatements and reclassifications. Now financial statements are restated if there is a material inaccuracy in a previous filing. Anyone can make mistakes when finding financial statements and this is of course accounted for in the rules. Now in practice setting up the right filters and warnings could save the day and this is after all a matter of specific or operational risk management. We can take action as soon as the correction becomes public information. We can even take action immediately for the most obvious mistakes. Now financial statements are also restated for a change in accounting rules when this change requires retrospective application. We have also restatements to account for mergers and spinoffs when the nature and size of the firm might change drastically in which case the firm should disclose a history that is coherent with its business and financial status post merger or post spin off. All restatements and disclosures are included in the 10QA, 10KA and 8KA filings. Now for discontinued activities both original statements and restatements are disclosed in the same filing. As it's related to discontinued activities we usually be reclassified as hell for sale. We will have reclassifications into change in net income from discontinued activities and cash flow from discontinued activities. Now a quick word on Europe. We have similar rules for European countries and firms reporting under IFRS. The policies convergence falls within the jurisdiction of the European Securities and Market Authorities the ESMA that was established in Jan 2011. Now ESMA is also working with the network of officially appointed mechanisms the OAMs for the central storage of regulated information. I have reported here some deadlines for France, Germany and Spain for the annual semi annual and quarterly filings. Now politics aside the harmonization takes time like pretty much anything in Europe. But if we take for the US case at least if we take everything we've mentioned so far this is how the final schedule looks like. Now 8K filings are followed by the filing of periodic reports the 10Q and 10Ks is a firm can't file before the deadline. It must notify the SEC through NT 10K and NT 10Q filings. I have pictured here two periods with the end of period EOP1 and EOP2. On the second period we might have restatements of previous filings and we also can have transition reports the 10QT and 10KT filings whenever there's a change in the fiscal year. Now imagine thousands of schedules like this one. One for each reporting firm with the end of periods do not necessarily match. Now that's what we can call a control disorder. So going back to our objectives and constraints we can fill in here the legal and regulatory constraints. I mentioned here systematic compliance to rules and regulations. Now I can hear your minds whispering. Why do we bother looking to the past? Why bother with restatements? After all I have the firm's guidance, I have a market consensus for earnings and we all know here that the market is all about expectations. Well I'd like to raise two questions. First question, how reliable is the firm's guidance for firms forward looking statements more specifically when it reflects the short-term horizon? It turns out there is an actual trade-off between short-term transparency that come with lower volatility, lower uncertainty and the drawbacks of the management that favors fast short-term profits over more profitable long-term projects. The pressure to disclose an exciting guidance for the short-term might come at the expense of a long-term performance. Now we've known for a while that there is no evidence whatsoever of any return upside to shareholders from issuing frequent earnings guidance. And this is why Google, Coca-Cola, Costco do not provide guidance more and more or reconsidering the benefits of providing guidance of short-term transparency in general as it makes these firms vulnerable to speculators. My second question is, how reliable is the market consensus when the analyst coverage is scarce? I should expect more opportunities on low coverage names. After all, high competition goes with low margin, low profit margin, sorry. So whenever competition is low, I must try to challenge the so-called consensus. Now overall, I would like an analysis that is free of any consensus bias, free of any limitations due to coverage. I need flexibility in treating new information so I can feel comfortable trading my portfolio not having to rely on black boxes. And to get there, I need to account for all potential result scenarios. I am hoping to assign a probability for each scenario given the firm's current financial health. If a scenario were realized in the past for a different firm but in the same industry and with a similar financial health then I should expect a higher probability to see this same scenario happen again. So I would like my model to learn from the history of comparable firms. In single terms, I need a cross-sectional sequential learning model, sorry, just one last word. This is why we need clean data. This is why we need a clean history and by clean, I mean coherence and I wish to avoid having for a firm both post-merger and pre-merger accounting data in the very same observation or using past revenues or predict future earnings failing to account for any discontinued business. Now after reading the game rules, it is never a bad idea just to watch a few rounds before taking a seat. So let's review here one case scenario which is the treatment, the balance sheet treatment to be exact of defined benefit pension plans. This is a recent change in accounting policies regarding the IFRS framework. One objective for this amendment is to align the balance sheet effects with those reported under US GAAP and part of the policies convergence. Now the treatment was revised in 2011 and the change was effective last year on Jan 2013. Of course what makes it interesting here is that this change requires retrospective application. I will quickly explain the basics of accounting. A plan is initiated by a contract offered by the firm to its employees. The firm records on its balance sheet a defined benefit liability or asset depending on the fund status of the plan. The defined benefit liability is the present value of the defined benefit obligations due by the firm towards its employees, mentioned here as DBO, less the fair value of the plan's assets. Now you can see here a synthesized pension plans balance sheet, the fund is legally separate from the firm and its sole purpose is to fund the employee benefits as specified by the contract. Now the present value of the DBO depends on various factors. I will just mention two here. The first one is the change in actual assumptions used to value the pension plans liabilities such as life expectancy or the discount rate. The discount rate is the discount rate used to discount the annuities. Now the discount rate is typically the implied interest rate of high quality corporate bonds. However, interest in assumptions can be revised and this revision will produce an actual gain or loss. Now there is also a dependency on plan amendments when moving from a defined benefit pension plan to a defined contribution plan, a dependency on plan curtailments when there is a significant drop in the number of employees covered by the plan and so on. Plan amendments and curtailments are defined as past service costs. Now these two effects can have very large impacts on the DBO and if they were directly recognized in the balance sheet in the firm's liabilities this could drive the balance sheet's volatility upwards. So a high volatility in the balance sheet is not something we're looking for. It could undermine the investor's trust in the reported figures. So before revisions firms were allowed to record these effects of balance sheets and the impact was subsequently and slowly recognized as an expense if their cumulative amount exceeded 10% of the plan assets. Now it turns out that many firms were actually deviating from the original purpose of this mechanism and recorded off balance sheets to delay their losses to equity and this way should have better financial leverage. Now the recycling methods and what is referred to as the 10% corridor methods is abolished with the new standards and since 2013 there is no amortization of actual losses or past service costs. These are immediately recognized in equity. Actual losses are recorded under a new item called remeasurements in other comprehensive income and past service costs are recognized in net income. Now take the Daimler case. Daimler has found the asset with a total value of 12 billion euros in 2012. Now the plan is underfunded. You can see here a negative 9.7 billion for the funded status. It had close to 8 billion of actual losses recorded off balance sheets. They were amortized slowly at the rate of roughly 150 million a year. In 2013 these actual losses were reclassified as remeasurements in other comprehensive income, net off taxes of deferred tax assets to be exact, which amounts to 6 billion euros. It's minus 8 and plus 2 over there. So 6 billion euros over 45.5 is roughly 13% of the total equity. Now furthermore the new standards removed any amortization of the actual loss that used to be recorded in net income. Now this had a positive impact on net profit, 29 minus 13 million for taxes, for a total of 16 million profit for the last quarter of 2012. Now obviously after revision, Daimler showed a slightly higher return on equity, a much lower comprehensive income, and a much higher financial leverage. Now if all ratios were affected equally across firms, my ranking would remain unchanged. Well it turns out that we have discrepancies across companies and more certainly across countries. For example, defined benefit pension plans are more frequent in France and Germany, and defined contribution plans are standard in the US and the UK. Regulations are also different. Germany allows for partial retirement agreements between management and employees when other European countries don't. So we've discussed the matter of data coherence between countries and firms. We have yet to discuss the matter of data coherence between market and accounting data. Well for example, if you have a second Republic offering, this second Republic offering will raise the equity value of a firm and the number of shares are sending at the same time. So in this case we would have a PE ratio, price to earnings ratio that is more or less unchanged, and the return on equity that should be lower. Now all these incoherences that I've mentioned so far might have dangerous implications for any systematic analysis. Whenever for example, training ratios are involved or for the calibration of a regression if momentum are used as regressors. Now if we get back to the Daimler case, we have our quarterly observations here piling up. We need to take four coherent consecutive quarters to annualize our flow data. Flow data consists here of net income. Income before a surgery item says cash flow from operations. Any item included in either the income statement or the cash flow statement. And these are opposed to stock variables such as total assets, shareholder's equity, all the items included on the balance sheets. Now we introduce here the concept of effective data. Now we mentioned that restatements are only available as of the beginning of 2013. So March 2013 will be my effective end of period. So for any information included in the amended filings, this information will be fully available as of March 2013. Now this means that all the lines highlighted in blue will be our four amended consecutive quarters. The four lines that will produce my adjusted data. Now the concept of effective dates will help mitigate any look ahead binds. It's very useful for simulations. I am only using here the information that I can see as of the reference dates that I choose, any reference dates in the past. Now let's play together and create our own exiting rules. I am looking to synthesize both market and accounting values of equity into one single variable, a score for each firm. The residual income model does so very nicely. The market and accounting information is synthesized into the equity cost of capital. Now the residual income model supports that the value of the present value of a company equals the present value of the current value of a company equals the present value of future excess earnings over the cost of equity capital. Now if you assume all future earnings to be value neutral starting as soon as next year, which means that the return on equity quickly declines to the cost of equity capital, then the cost of equity will exactly reflect the earnings to price ratio, the earnings yield ratio. It is exactly the inverted price to earnings ratio. Now Gebart in 2001 assumes that future excess earnings converge towards a neutral value represented by the industry mean. Now this convergence is assumed to be achieved in year 12. He predicts the one and two years ahead return on equity and assumes that this return on equity will fade slowly after the third year towards an industry mean. Now how far exactly should you predict the firm's earnings? Well the answer depends on the objective that we hope our model will achieve. And this objective is one or two years ahead forecast of the ranking of our stock performances. Lastly, if I am looking to produce a ranking only within the same industry, I don't really need the model to account for the concept of an industry mean. I don't need to complexify my model. Now this is not 100% true but this is another story. Now if we look at the earnings residual model through the eyes of Fama and French, this model emphasizes the mutual dependency between the firm's profitability represented by the return on equity, the firm's market value represented by the price-to-book ratio and the firm's propensity for reinvesting its earnings represented by the equity growth. And these dependencies are reflected by the cost of equity capital that is here, the required return on equity for a shareholder. Now the residual income model supports three relationships. First relationship, holding everything else constant, the required return on equity should increase with profitability. Second relationship, holding everything else constant, the required return on equity should decrease when market value increases. So one on top. And finally, the final relationship is based on the fact that earnings are either reinvested in the firm's project or distributed to shareholders. So increasing the level of reinvestments without any positive prospects on profitability nor market value for the shareholder should drive down the required return on equity. In this case, distributing dividends for the firm should be more, for the shareholder, it should be more interesting than reinvesting the earnings. Now we can observe the equity market value within regular trading hours. So we will need to predict the return on equity and the growth in the book equity. Now the forecasts are produced by, we mentioned this earlier, a cross-sectional model. This cross-sectional model uses as factors the earnings quality, profits distribution, revenue, size, operating and financial leverage for all the firms of our universe. Now we are going to simplify this model to the point that all we need are the one and two years ahead forecasts of the return on equity. I didn't mention it here, but we could use also macro and external factors in our cross-sectional model. So simplifying the setups suggested by Gebart, Fama and French is a luxury that we can afford because we are adopting a relative value approach. If I am looking to compare firms within the same industry, I need only discriminant factors within this industry. So I choose here to ignore all macroeconomic factors, all external factors, I assume will have a similar impact over the firms of my industry. These firms will belong to the same sector and will abide to the rules and regulations of the same country, at least when it comes to reporting requirements. Now these are strong assumptions, but still reasonable given what we are expecting from this model, which is an apples-to-apples relative value analysis for the year ahead. Furthermore, remember that we assumed all earnings to be either reinvested or distributed to shareholders. Now the clean surplus hypothesis, holds that the growth of book equity is completely explained by the earnings and the dividends distributed. Now if you assume like Gebart that the dividend payout ratio is constant in the future, then we have the growth of equity that is completely determined by the one and two years ahead return on equity. So as promised, this is what we have left. One year ahead and two years ahead return on equity to a forecast. Now a word on the calibration. The timing of observations matters if we are considering panel models. If periods covered by the income and cash flow data were assumed to be exactly the same, perfectly aligned quarter to quarter, then panel models would be ideal. Now the major advantage of panel models lies in the detection and treatment of autocorrelations in the editor. Basically we can theoretically detect whenever the model is better specified as a moving average. Now panel models allow also for the treatment of seasonality. Now the perfect alignment of observations could be achieved by altering the original data using data interpolation or the so-called calendarization techniques. Now the calendarization process is a simple idea to begin with. I have my end of periods in February, May, August and November every quarter and I would like to produce data estimates for the standard quarters ending in March, June, September and December. Now the thing is working with estimates instead of actual data is never an easy solution. I'd like to avoid that here. So I will concentrate my efforts on the simple cross-sectional model and use momentum regressors to remove potential autocorrelations in their return. Now basically you can refer to the first chapters of full-reach, a book entitled Econometrics of Cross-sectional and Panel Data. You'll find there a discussion of the benefits of momentum in improving the model specifications. By the way, a full bibliography is available at the end of the documents. Now Fama and French used annual data from the 10K findings from beginning of July to end of June to calibrate the model and predicts one, two and three years ahead returns on equity. You can see here predicts T plus one, T plus two, T plus three forming an observation. Our calibration, on the other hand, is performed on every results announcement date or a few business days later. Now whenever new information is acquired we are taking as a reference the last end of periods available and moving backwards this is the last end of periods and we're moving backwards to collect all data and again the predictive T plus one will form one observation and so on and these observations are used to calibrate our cross-sectional model. Now every newly acquired observation for one single firm will impact the calibration. So if it impacts the calibration it will also impact our forecasts for all the other firms of our industry. This is the sequential learning property of the model. Now once I have my estimates for the cost of capital I take an active long position on the highest and a short active position on the lowest ICCs. ICC is meant for implied cost of capital here. I choose to test here the ranking quality for 40 long positions and 40 short positions against the realized returns and to do so I have rank correlations. One classic ranking correlation is the Spearman rule the first one mentioned here. We also have the Kennel Tower. Basically I'm assigning to the same to the 41st ICCs the same value of 40 plus one over two which is 20.5 and I assign the same rank on the 40 highest ICCs. 2 times n minus 39 over 2. The highest correlation is reached 57% of the time within the first three months from the estimation. You can see here this is the daily average correlations that are at a peak when volatility is high in 2008 and here 2011. The maximum correlation one more thing this is the level of significance for a single correlation which is around 10%. It depends on the size of your sample. Here I have this is a technology sector so I have roughly 200 firms in my sample after filtering. This is the maximum correlation which is reached very often within the first 25 days at 15 to 20%. That's the correlation that we are looking to buy here. The strategy is mean reverting so the higher the correlation the better our hope to generate alpha. We are long correlations. We are done. We still need market risk management framework. Here I have completed the risk objective. I mentioned small beta allocation diversification and specific risk control. The terms small beta in small beta allocation cover a wide range of procedures. Every problem that accounts for a measure of risk of high variance volatility and correlation qualifies for many as small beta. I would have liked to catch your attention on some of the main challenges here for the allocation but I will have to pass because we are a little bit short of time but you can see here on this slide the back-testing results for the technology sector in dashed blue index. The NDXT is in dashed red and in light blue, dark blue and green we have the distribution of the cost of equity capital. We can see in 2008 that the cost of capital have reacted with months of delay to the crisis. Again we cannot expect the model to focus macroeconomic events but we can hope like in 2008 and 2011 that it would capture and profit from high volatility periods. Whenever we see high discrimination between stocks. Now I have reported here also this is the last slide a sector neutral simulation on all US names. It's a reachable target, it would be interesting to see how a long-only version with same sector weights as a classic benchmark performs. The source Hino is decent and the draw down is here to remind us that we are dealing with stocks so always highly risky. It happens here in the crisis. These are distributions per sector liquidity and capitalization I will let you analyze that on your own. Thank you very much, unfortunately time is always too short and I hope this presentation gave you at least an idea of the challenges of systematic management for equity shares. Are we now answering any questions that you might have? Any questions? Yes? Yes, it turns out recently recently Yes, basically she's asking how we could retrieve the financial information from statements knowing that basically it's an exhaustive process. It seems like an exhaustive process and it is actually, but the thing is recently the SEC asked for the firms that were to report their findings to add an XBRL file XBRL extensive business so basically you can do it, it's possible, it's a lot of work I've worked on it, but it's a lot of work and you can actually work on some mappings for any given firm you will have mostly the same items that we find in the statements on a smart mapping you can directly retrieve this information otherwise you have providers like Edgar, but of course they will give you the information once you have the 10Ks or 10Q reports there is always because the number of information we need the number of variables we need is very limited especially for this model I need a dozen of variables so I can maximum add them manually once I have my 8K I hope this is satisfying any other question? I know it seems extremely theoretical and can be very complicated for people who didn't go into statistical models I mean I hope I gave you at least an idea of what we can achieve here everything that I mentioned is really inspired by the work of Fama in French and the work of Gebart initially but I pushed it towards relative value it's not about getting an absolute level for the equity premium but getting a relative value with respect to the same stocks of an homogeneous universe, that's the idea well I think this is very much CFA stuff so if that's not something that you would like to do let's give you a round of applause I was particularly pleased that you mentioned XPRL as well the Extensible Business Reporting Language under President of the Swiss XPRL through Restriction so that's one thing that we should definitely push for even though XPRL in Switzerland is very much an actor issue due to the low-key interest of issuers who just see it as an additional obligation and the Stock Exchange also is not really very keen on pushing for that because of that because Stock Exchange as you all know is a member of the association and if some of their members actually push against that and see it as an obligation they will also be somewhat reluctant to adopt it so thanks thank you very much thank you