copyright 2003-2010 uk-neural.net Passing the Bank Manager test Not too many years ago, anyone wanting a loan or mortgage would face a formal interview with a local branch manager. After his interrogation, a check of your current financial status and sponsors/guarantors, the manager pronounced if you would, or would not, be offered the facility. A decision which was often declared after some days delay, and by post. Some applicants, deemed unsuitable by the counter clerk at first enquiry, would not have even reached the interview stage. Today's applicants are dealt with by general bank staff. A form is filled and a decision is given within minutes, the money can be in your account shortly afterwards. Customers may prefer the old or the new ways, but from a business point of view the latter reigns supreme. It is far less costly to process and the automated version makes much better decisions than your average pin-striped bank manager ever did. The difference in procedures is striking, and the change is possible because of data driven machine intelligence. Banks, like all other institutions, keep records of customers, transactions and outcomes. This gold mine of information can be interrogated to reveal many things, but here it can be used to collate a list of past loan applicants, and the outcome of each particular agreement. With thousands of these examples to work with artificial intelligence software can "learn", so that when a new set of conditions is presented the computer can suggest whether this would be a good or bad risk. It could be argued that anyone with a modicum of common sense could do the same. This is true, but only up to a point, maybe the very best and worst applicants could be identified this way. But, whilst selecting only the very best candidates would keep defaulters to a minimum, turning so much business away would directly impact on turnover. The best business models need to strike the best balance between risk and reward, the trick here is not to eliminate defaulters entirely, but to know what percentage of bad debt the business model can accommodate for the best return. Traditional statistics can often fail in this type of number crunching because there is no definite good or bad, right or wrong, black or white, strong or weak. Even the very best customer whose profile scores the highest in every category can default. Manchester United do not win every game, Tiger Woods does not win every tournament. If information is power, then the power we seek here is the probability of x, y or z. Machine learning software is ideally suited to such tasks. Establishing and grading preconditions relevant to the applicant in place at the time a previous loan was granted, the final outcomes can be marked as '1' for the perfect customer through to '0' for the customer who defaults within the first year. Once the artificial intelligence engine has processed these data, the same precondition details for any new applicant can be fed in and a "credit risk" prediction generated. From then on, by feeding the computer with details of a new applicant's age, qualifications, employment history, current employment, salary, credit history, marital status, number of children, etc., etc., fast processing in the local branch will quickly credit score the application in the range of 0 to 1. Specific business conditions will determine where the yes/no line might be drawn as a result. For example, scores below 0.2 rejected, above 0.75 accepted and offered the very best terms. Those in the range 0.2 - 0.75 accepted, but at slightly higher interest rate and/or a shorter repayment period. Machine learning is not limited to credit risk assessments. Indeed anyone with access to records of past events, together with what could be thought of as influential related conditions in place prior to the outcome has the basic building blocks to predict what may happen given a fresh set of circumstances. Here's a couple of case study overviews to help give you ideas for how your business could benefit. The role of telephone call centre operative is not one that suits everyone and selecting the wrong candidate is a costly exercise. Dave S, the Section Manager had tried almost everything to select new recruits with the best potential at the interview stage, but the staff turnover rate remained stubbornly high. One evening, over a drink with a friend neural networks were suggested. Dave collated a standardised 'profile' of past applicants who'd been appointed - age, sex, last job & salary, education, in a relationship, children, etc., etc. and for each of these allocated a score of 0 to 100. A 100 mark for those who'd gone on to be enthusiastic and inspirational team leaders, with a zero reserved for the girl who quit before the end of her first day! Neural network software was trained using these records. A revised application form was created and duly sent to every new job enquirer with the now standard profile question in place. Once returned, each set of answers were then individually processed using the specifically trained net. As a result of this simple network application Dave was forearmed as never before when screening potential employees. His people skills and intuition were still needed, but armed with the emotion-free conclusions output from his neural net, his interview decisions proved to be far better than ever before. The suitability of new recruits improved measurably, staff turnover dramatically reduced and the short-stay new starters were a thing of the past. Dave's contribution was recognised by a promotion, salary increase and has been charged with implementing a recruitment policy throughout the group. A well established major car brand franchise and used car sales business wanted to make their newspaper advertising budget more effective. From past records they could access individual advertisements and match the two-week sales figures that followed. Cause & Effect? Their local paper ads could be categorised several ways; There are three newspapers covering the target areas, each of which may be used or not. (2 x 2 x 2 = 8 possibilities) Full colour, spot colour or black & white Double-page spread (either centre pull-out of newspaper or elsewhere) Full page (right or left hand page) Half page (right/left and top/bottom of page) Quarter page (right/left and top right, top left, bottom right, bottom left) Loose flyer enclosed with newspaper (A5 single sided, A5 double sided, A4 unfolded, A4 folded) Advertising content can be categorised as (a) Top featuring a specific new car model, (b) Equal billing to more than one model (c) Featuring the entire new car range or (d) Emphasis on the used car stock. Occasionally combinations of (a&d), (b&d) or (c&d) A pause here to allow you to appreciate how quickly the possible combinations grow. Already, from first point down we have 8 x 3 x 2 x 2 x 4 x 8 x 7 possibilities, allowing for over twenty-one thousand variations. The month of the year is also very influential in car sales figures too. Factoring in these 12 further possibilities grows our variations here to close on a quarter of a million. Factor in business side effects like a new model launch, model facelifts, money off deals, minimum trade-in deals, etc. and you'll quickly realise that trying each variation in an attempt to pick the best mix is impossible. Cue machine learning software. Train a net on what has happened Then rank new proposed campaigns by processing each through the specially built software model. The software will suggest the potential profitability of each. The machine learning strategy. Efficient, effective and productive aid to decision making. Impress your boss! advertiser space uk-neural.net  anticipating the future by examining the past