Function Point Prognosis © by Manfred Bundschuh
Function Point Prognosis Revisited
Regression Analysis for Approximation of Function Point Counts
Dipl.-Math. Manfred Bundschuh, CFPS
AXA Colonia, Cologne, President of DASMA
Contents:
Function Point Prognosis Revisited 1
Regression Analysis for Approximation of Function Point Counts 1
1 Environment 1
2 Project Register Data Base 1
3 First Results 2
3.1 Function Proportions 2
3.2 Average Function Complexity 3
3.3 Function Ratios 3
3.4 Correlations 4
4 Main Result 5
5 Error Discussion 7
6 Conclusion 7
7 Literature 7
8 Figures page 7
1 Environment
From 1996 to 1998 about 40 Functon Point counts with totally 77,865 Function Points ( FP’s) have been accomplished in CNV AG, the outsourced non-insurance part of AXA Colonia Insurance. FP’s mean througout this paper IFPUG 4.0 unadjusted FP’s. Function Point counts are obligatory at least at the end of the Requirements Analysis, end of the Design Phase and at project post mortem. Small projects are only counted twice. (not at the end of the Design Phase). Function Point Prognosis instead of count is obligatory during the Feasibility study and at Project start.
The IT Department of the CNV AG includes about 500 IT Professionals with approximately 50 project leaders. The insurance branches deliver about 160 IT coordinators supporting the IT projects. IT development is mostly host based with COBOL / Generator programming. There are about 500 IMS databases and 2,000 DB2 tables, 1.6 Million IMS- and 1.2 Million CICS transactions per day. PC projects use Optima++, a C++ Shell, for programming. Projects develop mostly interactive database administration systems for e.g. car or life insurance or claims management, within a very complex environment with centralized data bases for e.g. insurance partners, cash systems, document management etc.
2 Project Register Data Base
Fig. 1 (see end of the report) shows detailed informations for each FP count as being kept in our project register, giving the quantity of EI’s, EO’s, EQ’s, ILF’s and EIF’s. If a project was counted repeatedly only the most actual count is shown and older counts are kept in a history table. There are sums of the quantity of EI’s and EO’s, of ILF’s and EIF’s which were needed for our research.
IO means througout this paper the sum of EI and EO.
The idea for this research was to find out if there were hidden informations in that data collection. The results proved to be very productive. In my first presentation[1] on the FESMA Conference 1998 in Antwerp I presented the results based on about 20 counts which increased to 40 counts today. Hence I was curious if the new results would be consistent and if parts of the project data base would deliver confident findings.
Fig. 2 ( see end of the report ) shows the according FP’s as well as sums for the FP’s of the EI’s, EO’s, EQ’s, ILF’s and EIF’s. The last two lines in this Table are for comparison with Morris/ Desharnais[2] and the Quick Estimate mode of Checkpoint, Estimation tool from Capers Jones‘ firm SPR (Software Productivity Research) in Burlington, MA.
3 First Results
3.1 Function Proportions
From Fig. 3 can easily be seen, that EO’s dominate in CNV AG ( 39% ) compared with the other two sources (22%-24%) in literature, whereas ILF’s are of minor importance (17% vs. 24%, 43%, resp. ). Because of this peculiarity one conclusion was not to use Checkpoint ( EIF + ILF = 46% vs. 23 % in CNV ) in Quick Estimate mode to estimate FP’s. The reason for the minor importance of ILF’s may be that CNV AG has many centralized databases which are widely refereced by almost all applications.
The domination of the EI’S and EO’s (together 64 %) seems to be the reason for the strong correlation between IO’s and the unadjusted FP’s- the main result of this research.
Function Proportions (Percentage) / Number of Projects / EI / EO / EQ / ILF / EIFCNV 1996 / 8 / 34 / 35 / 11 / 18 / 2
CNV 1997 / 12 / 18 / 43 / 12 / 18 / 9
CNV 1996/97
= FESMA 1998 / 20 / 27 / 39 / 11 / 18 / 5
CNV 1998 / 19 / 21 / 41 / 13 / 15 / 9
CNV S (1996-1998) / 39 / 25 / 39 / 14 / 17 / 6
ISBSG Rel. 5 / 451 / 37,2 / 23,5 / 13,2 / 22,3 / 3,8
Metricviews / 26-39 / 22-24 / 12-14 / 24 / 4-12
Checkpoint / 20 / 24 / 10 / 43 / 3
Nigel Scrivens *) / 33 / 21 / 19 / 23 / 4
ISBSG Rel. 5
Enhancement Projects / 119 / 36 / 32 / 12 / 15 / 5
Average of this table: / 28,6 / 26 / 13,8 / 20,1 / 5,8
*) Nigel Scrivens from shl.com via Internet reported industry averages
Fig. 3: Function Ratios
Fig. 3 shows that there are only small differences in the ratios of the functions between the two investigations. Hence this result proved to be steady since my last presentation with only half as much counts. Of course I left one runaway (Project No. 21 in Fig. 1) out of the calculation. It is our centralized data base for document management which delivered 38,863 EO FP’s in a count of 39,044 unadjusted FP’s.
The results of Fig. 4 can be used as a rule of thumb, if either the EI’s, EO’s, EQ’s, ILF’s or EIF’s of a count are known (e.g. FP’s = 100 * EI / 28,6 using the last table row), thus leading to a „FP - Prognosis“.
3.2 Average Function Complexity
Another one of the first results of this data collection was, that the VAF (Value Adjustment Factor) in CNV AG is typically between 0.7 and 1.2 ( Host average 0,94 and PC average 0,96, both average 0.95 ) and about 0.7 for migrations.
Fig. 1 and 2 were analysed with the EXCEL problem solver in order to find out how many FP’s a „typical“ EI, EO, EQ, ILF or EIF has in our environment. The results are shown in Fig. 4, the values are given for PC - Projects, Host - Projects and both.
AverageFunction Complexity / EI / EO / EQ / ILF / EIF
IFPUG / 4 / 5 / 4 / 10 / 7
ISBSG
Rel. 5 / Total / 4,3 / 5,4 / 3,8 / 7,4 / 5,5
Asia Pacific / 4,0 / 5,6 / 3,9 / 7,4 / 5,6
Europe / 4,2 / 4,9 / 3,8 / 7,2 / 5,3
North America / 4,9 / 5,2 / 3,8 / 7,6 / 5,5
CNV 1998 / All / 4,6 / 5,5 / 4,3 / 8,0 / 5,9
PC / 4,1 / 5,7 / 3,9 / 7,1 / 5,3
Host / 4,9 / 5,5 / 4,6 / 8,5 / 6,1
CNV 1999 / All / 4,6 / 5,7 / 4,3 / 8,2 / 6,1
PC / 4,0 / 5,7 / 3,9 / 7,3 / 5,4
Host / 4,8 / 5,7 / 4,5 / 8,5 / 6,2
Fig. 4: Average Function Complexity
SPR Function Points use the average IFPUG classification for Function Point estimation. Instead of SPR Function Points we advise our project leaders to use our „CNV Typical average FP’s classification“.
The results of Fig. 4 can be used as a rule of thumb, if the EI’s, EO’s, EQ’s, ILF’s and EIF’s of a count are not yet qualified as low, average or high, but are known in quantity, thus leading to a „FP - Prognosis“. Naturally, both Fig. 3 and 4 can also be combined for the purpose of an early Prognosis.
3.3 Function Ratios
Charley Tichenor from IRS Information Systems found for IRS’s the tax processing (batch) Software a similiar strong correlation as we did with R2 = 0.9483 for the prognosis of unadjusted Function Points counts from ILF’s, instead of our IO’s. Fig. 5 shows the differences between IRS ( typical batch environment ) and CNV ( typical online environment ).
Fig. 5: Comparison with Tichenors ILF model
One would expect three inputs (add, change, delete), one output and one enquiry for maintenance of a file. Fig. 6 shows the ISBSG Rel. 5 averages together with the CNV and Tichenor’s results.
Function Ratios / Number of Projects / EI : ILF / EO : ILF / EQ : ILF / EIF : ILFTichenor[3] / 0,3331 / 0,3903 / 0,011 / 0,097
CNV / 1996 / 8 / 3,27 / 2,75 / 1,10 / 0,14
1997 / 12 / 1,89 / 3,65 / 1,35 / 0,75
S 1996/97 = FESMA 1998 / 20 / 2,66 / 3,14 / 1,20 / 0,4
1998 / 19 / 2,63 / 3,87 / 1,79 / 0,82
S 1996 - 1998 / 39 / 2,66 / 3,33 / 1,36 / 0,51
Total / 238 / 2,9 / 1,5 / 1,1
ISBSG / Asia Pacific / 116 / 2,6 / 1,1 / 0,9
Rel. 5 / Europe / 32 / 3,8 / 2,6 / 1,9
North America / 90 / 0,9 / 1,9 / 1,3
Fig. 6: Comparison of Function Ratios
3.4 Correlations
Fig. 1 and 2 were also analysed with the EXCEL problem solver in order to find correlations between the quantities of the EI’s, EO’s, EQ’s, ILF’s or EIF’s and the total Number of FP’s for each count (the main result of this paper). For this reason regression analysis was performed. The correlation coefficients R2 are shown in Fig. 7. Results with R2 ³ 0.95 were accepted as relevant.
R2 / 1999 / 1998Total / 0.9509 / 0,95
PC / 0.9580
Host / 0,9760
EI / 0,66 / 0,77
EO / 0,82 / 0,84
Host EI / 0,77
Host EO / 0,91
> 1,500 FP’s (1998 > 1,800) / 0,61 / 0,71
< 1,500 FP’s (1998 < 1,200) / 0,93 / 0,95
Fig. 7: Determination coefficients
It can easily be seen that the usage of FP’s instead of the quantity for the Prognosis gives a stronger correlation, but the higher effort for counting is not adequate to the higher precision. One should always keep in mind that estimation has to do with uncertainty per se.
4 Main Result
The result of the research was, that the sum of the quantities of EI’s and EO’s ( IO’s in our diction ) is correlated with R2 ³ 0.95 (R ³ 0.97) to the total amount of FP’s of a count and can thus be used as a „rule of thumb“ for the „Prognosis“ of FP’s when the FP’s of EQ’s , ILF’s and EIF’s are not known.
Fig. 8: Regression Analysis of Main Result
Fig. 9 shows the formulae for our CNV FP- Prognosis.
CNV Prognosis / Formula / R21998 / FP = 7,3 IO + 56 / 0,9525
1999 / Total / FP = 7,6 IO + 39 / 0,9509
PC / FP = 6,5 IO + 134 / 0,9760
Host / FP = 7,8 IO + 11 / 0,9580
Fig. 9: FP Prognosis Formulae
5 Error Discussion
Last year we tested the result achieved by using only the quantity of EI’s an EO’s ( summarized, =IO’s in out diction ) for each count, computing it with the regression formula and determing the difference between this computation and the counted FP’s finding an error range of 12.11% ( Median 10.6% ). Unfortunately this year there was no capacity to redo this examination with the enhanced data set..
6 Conclusion
Thus we recommend our project leaders our rule of thumb for FP - Prognosis ( main result ) with an error range of 15 - 20 % - but only for early FP - Prognosis. Of course a complete FP count at the end of the requirements analysis is obligatory. Astonishingly our early findings of last year could be proved valuable and only small differences showed up.
Our findings compared with those from other firms show that such data collections can be used to find heuristic solutions for FP - Prognosis, either using Function Point Proportions („typical FP’s“), Function Ratios or regression formulae. But there is evidence that different environments demand for according solutions. Hence each firm should develop its own know - body of heuristic solutions and should distingnish between different development platforms etc. when doing so.
In 1998 half of our applications in CNV AG were counted and base counts still go on (only a few this year due to Y2K, EURO and Migration of the Albingia Insurance into the AXA Colonia. Meanwhile each IT group has at least one trained FP Counter. Hence we hope to get more reliable data and will keep this rule of thumb up to date. We will not count but estimate the FP’S of SAP, using this prognosis.
7 Literature
1. M. Bundschuh, Function Point Prognosis, FESMA 98 „Business Improvement through Software
Measurement“, Antwerp, Belgium, May 6-8, 1998, pp. 463 - 472
2. M. Morris, J. M. Desharnais:
„Validation of the Function Point Counts“, in: Metricviews, Summer 1996, p. 30
3. Charley Tichenor, from IRS in Proceedings of the IFPUG Fall conference „Target 2000“, Scottsdale, Arizona, USA, Sep. 15 - 19, 1997, pp 299 - 321
4. John E. Gaffney, Jr. from Corp. Software Productivity Consortium in Herudon, Virginia 22070 in his Presentation „A Simplified Function Point Measure“ at the IFPUG 1994 Fall Conference, Oct. 19-21, 1994 in Salt Lake City, Utah
8 Figures page
Fig. 1 FP Statistics, FPM - Quantities...... 7
Fig. 2 FP Statistics, FPM - Function Points...... 8
Fig. 3 fUNCTION RATIOS...... 2
Fig. 4 AVERAGE Function COMPLEXITY...... 3
Fig. 5: Comparison with Tichenors ILF model...... 4
Fig. 6 COMPARISON OF FUNCTION RATIOS...... 4
Fig. 7 Determination coefficients...... 3
Fig. 8 Regression analysis of main result...... 4
Fig. 9 FP - Prognosis formulaE...... 10
page 8 of 8
Function Point Prognosis © by Manfred Bundschuh
CNV - Fig. 1 Statistics
page 8 of 8
[1] M. Bundschuh, Function Point Prognosis, FESMA 98 „Business Improvement through Software Measurement“, Antwerp, Belgium, May 6-8, 1998, pp. 463 - 472
[2] P. M. Morris (Total Metrics), J. M. Desharnais (SELAM): „Validation of the Function Point Counts“, in: Metricviews, Summer 1996, p. 30
[3] Charly Tichenor: „The IRS Development and Application of the Internal LOGICAL FILE Model to estimate Function Point counts“, in his presentation at he IFPUG Fall conference „Target 2000“ in Scottsdale, Arizona, USA, Sep. 15 - 19, 1997, pp. 299 - 321,