λ Building Fintech App in Java Tips & Trick 

THE SPEAKER › Alex Xandra Albert Sim › Principal R&D Engineer at Blibli.com › @bertzzie(.sim) 

λ Disclaimer Presentations are intended for educational purposes only and not to replace independent professional judgment. The views and opinions expressed in this presentation do not necessarily reflect the official policy or position of blibli.com. Audience discretion is advised. 

λ Fundamentals of Financial Computation 

λ Foundational Knowledge › Computational Finance (Keuangan Komputasional) › Emphasis on: – Mathematical Finance (Matematika Keuangan; Teknik Finansial) – Numerical Method (Metode Numerik) › Layman terms: – Analysis – Simulation – Approximation 

λ Practical Concerns › Calculations in fintech needs to be accurate and precise › Calculations needs to be FAST 

λ How? 

λ Accuracy and Precision Java Fintech Tip & Trick #1 

λ #1 Floating-point Numbers are FAKE NEWS 

λ Let’s look at some code Click for code 

λ Output 

λ Notice › When the fraction is power of 2 (1/2, 1/4) we got correct result › 1/5 is correct before being summed › 1/3, 1/6, and 1/7 have tiny roundoff-error › 1/3 has round-up error on fraction, but round-down error on sum. Vice versa for 1/6 and 1/7 

λ But HOW? 

λ The Float Number Format Sign Exponent Fraction Sign: positive (0) or negative (1) Exponent: 127 or -126 to +127 Fraction: the actual fraction value with leading 1 for positive (normalized) and leading 0 for 0 and others (denormalized) -5.5 = 1 1000 0001 011 0000 0000 0000 0000 0000 

λ The Float Number Format Sign Exponent Fraction Normalized value: = −1 × 2 × 1. Denormalized value: = −1 × 2−126 × 0. 

λ The Float Number Format Sign Exponent Fraction -5.5 = 1 1000 0001 011 0000 0000 0000 0000 0000 011 0000 0000 0000 0000 0000 1 + 0 + 1 21 + 1 22 + 0 + 0 + 0 + ⋯ 0 1 1 0 0 0 … 

λ Example! Let’s do some math! 

λ Conclusion › Float (and double) should not be used to count money (unless you’re counting my debt – then by all means) › They are still useful because it’s fast and quite accurate. And sometimes the mistakes are funny. 

λ Conclusion (cont.) Use Float (and Double) with Discretion! When in doubt, use BigDecimal 

λ Overflow Java Fintech Tip & Trick #2 

λ #2 Watch out for Integer Overflows! 

λ Java’s Integer Integer Type Size (bits) Min Value Max Value byte 8 -128 127 short 16 -32.768 32.767 char 16 0 65.535 int 32 -2.147.483.648 2.147.483.647 long 64 -9.223.372.036.854.775.808 9.223.372.036.854.775.807 

λ The Fly-Trap › Notice how max value is always smaller than the positive of min-value? What’s the output? -2147483648 -2147483648 

λ The Fly-Trap (2) › Notice how max value is always smaller than the positive of min-value? What’s the output? -2147483648 -2147483648 

λ The Fly-Trap (3) › Notice how max value is always smaller than the positive of min-value? What’s the output? 2147483640 + 10000 = -2147473656 

λ More Unsafe Code Example 

λ Safe Code (Java 8+) There’s also Math.subtractExact and Math.negateExact 

λ Safe Code (Bound Check) 

λ Safe Code (Bound Check) (2) 

λ Safe Code (Bound Check) (3) How to do bound check for abs? 

λ Safe Code (Casting) 

λ Safe Code Simply: Use BigInteger 

λ As Binary As Possible Java Fintech Tip & Trick #3 

λ Binary › Everything inside our computer is represented by binary › There’s usually a way to efficiently calculate stuff with binary, since both CPU and Memory works in Binary 

λ Binary Example: Power › Simple example, to calculate: = › The simple and naïve approach: = ∙ ∙ ∙ ∙ … › Finishes in linear time: 

λ Power: the Better Approach = = 14 Get binary rep. of 1110 Convert to power of 2 (right to left) 1 ∙ 2 ∙ 4 ∙ 8 Remove the 0 bit 2 ∙ 4 ∙ 8 Calculate! 

λ Power Comparison! Click for code 

λ Power Comparison Result! Result "tech.namas.sharing.numbers.PowerComparison.powBinary": 1188449.776 ±(99.9%) 133027.111 ops/ms [Average] (min, avg, max) = (938259.958, 1188449.776, 1379370.603), stdev = 153194.219 CI (99.9%): [1055422.665, 1321476.887] (assumes normal distribution) Result "tech.namas.sharing.numbers.PowerComparison.powMath": 11479.744 ±(99.9%) 429.096 ops/ms [Average] (min, avg, max) = (9987.543, 11479.744, 12010.515), stdev = 494.148 CI (99.9%): [11050.648, 11908.840] (assumes normal distribution) Benchmark Mode Cnt Score Error Units PowerComparison.powBinary thrpt 20 1188449.776 ± 133027.111 ops/ms PowerComparison.powMath thrpt 20 11479.744 ± 429.096 ops/ms 

λ Benchmarking Questions › ~10x difference? Is this real life? Or is this fantasy? › Why do we need the (min, max, avg) numbers? › Why is the std-dev differs so much? › What does “normal distribution means? 

λ Interpolation & Approximation Java Fintech Tip & Trick #4 

λ Prediction › A lot of fintech use case is analyzing / predicting future values › Will person x, given his/her history, pay his/her debt in time and consistently? › Will investing in share A yield good value given its share value history? 

λ Interpolation 

λ Interpolation (n) /in-​ˌtər-pə-​ˈlā-shən/ a method of constructing new data points within the range of a discrete set of known data points Approximation (n) /əˌpräksəˈmāSH(ə)n/ a process of matching two functions that closely match, usually from the same class for specific task 

λ The Common Polynomial Form › Most function to calculate stuff like this will be shown in polynomial form: = 0 + 1 + 2 2 + ⋯ + › Which are very expensive to calculate (look at all the Math.pow) 

λ The Newton Form › To simplify, we could use the Newton Form: = 0 + 1 − 0 + 2 − 0 − 1 + … + − 0 − 1 … − −1 › which allows us to cache calculations and apply dynamic programming techniques 

λ Polynomial to Newton form 

λ Let’s look at some code Click for code 

λ Final Result 

λ Closing 

λ Summary › Don’t use float. It’s not accurate and it adds up. › Beware of overflows, specifically for big calculations. › Optimize by exploiting the fact that our computer uses binary internally. › Approximate and interpolate where possible. 

λ Further Discussion › How to store transactions and values? › How to do reporting / analytics fast and real-time? › How to create instant decision-support? 

λ Question and Answer